| 1 | #include <unistd.h> |
| 2 | #include <fcntl.h> |
| 3 | #include <string.h> |
| 4 | #include <signal.h> |
| 5 | #include <time.h> |
| 6 | #include <assert.h> |
| 7 | |
| 8 | #include "fio.h" |
| 9 | #include "hash.h" |
| 10 | |
| 11 | /* |
| 12 | * Change this define to play with the timeout handling |
| 13 | */ |
| 14 | #undef FIO_USE_TIMEOUT |
| 15 | |
| 16 | struct io_completion_data { |
| 17 | int nr; /* input */ |
| 18 | |
| 19 | int error; /* output */ |
| 20 | unsigned long bytes_done[2]; /* output */ |
| 21 | struct timeval time; /* output */ |
| 22 | }; |
| 23 | |
| 24 | /* |
| 25 | * The ->file_map[] contains a map of blocks we have or have not done io |
| 26 | * to yet. Used to make sure we cover the entire range in a fair fashion. |
| 27 | */ |
| 28 | static int random_map_free(struct fio_file *f, const unsigned long long block) |
| 29 | { |
| 30 | unsigned int idx = RAND_MAP_IDX(f, block); |
| 31 | unsigned int bit = RAND_MAP_BIT(f, block); |
| 32 | |
| 33 | dprint(FD_RANDOM, "free: b=%llu, idx=%u, bit=%u\n", block, idx, bit); |
| 34 | |
| 35 | return (f->file_map[idx] & (1 << bit)) == 0; |
| 36 | } |
| 37 | |
| 38 | /* |
| 39 | * Mark a given offset as used in the map. |
| 40 | */ |
| 41 | static void mark_random_map(struct thread_data *td, struct io_u *io_u) |
| 42 | { |
| 43 | unsigned int min_bs = td->o.rw_min_bs; |
| 44 | struct fio_file *f = io_u->file; |
| 45 | unsigned long long block; |
| 46 | unsigned int blocks, nr_blocks; |
| 47 | |
| 48 | block = (io_u->offset - f->file_offset) / (unsigned long long) min_bs; |
| 49 | nr_blocks = (io_u->buflen + min_bs - 1) / min_bs; |
| 50 | blocks = 0; |
| 51 | |
| 52 | while (nr_blocks) { |
| 53 | unsigned int this_blocks, mask; |
| 54 | unsigned int idx, bit; |
| 55 | |
| 56 | /* |
| 57 | * If we have a mixed random workload, we may |
| 58 | * encounter blocks we already did IO to. |
| 59 | */ |
| 60 | if ((td->o.ddir_nr == 1) && !random_map_free(f, block)) { |
| 61 | if (!blocks) |
| 62 | blocks = 1; |
| 63 | break; |
| 64 | } |
| 65 | |
| 66 | idx = RAND_MAP_IDX(f, block); |
| 67 | bit = RAND_MAP_BIT(f, block); |
| 68 | |
| 69 | fio_assert(td, idx < f->num_maps); |
| 70 | |
| 71 | this_blocks = nr_blocks; |
| 72 | if (this_blocks + bit > BLOCKS_PER_MAP) |
| 73 | this_blocks = BLOCKS_PER_MAP - bit; |
| 74 | |
| 75 | if (this_blocks == BLOCKS_PER_MAP) |
| 76 | mask = -1U; |
| 77 | else |
| 78 | mask = ((1U << this_blocks) - 1) << bit; |
| 79 | |
| 80 | fio_assert(td, !(f->file_map[idx] & mask)); |
| 81 | f->file_map[idx] |= mask; |
| 82 | nr_blocks -= this_blocks; |
| 83 | blocks += this_blocks; |
| 84 | block += this_blocks; |
| 85 | } |
| 86 | |
| 87 | if ((blocks * min_bs) < io_u->buflen) |
| 88 | io_u->buflen = blocks * min_bs; |
| 89 | } |
| 90 | |
| 91 | static unsigned long long last_block(struct thread_data *td, struct fio_file *f, |
| 92 | enum fio_ddir ddir) |
| 93 | { |
| 94 | unsigned long long max_blocks; |
| 95 | unsigned long long max_size; |
| 96 | |
| 97 | /* |
| 98 | * Hmm, should we make sure that ->io_size <= ->real_file_size? |
| 99 | */ |
| 100 | max_size = f->io_size; |
| 101 | if (max_size > f->real_file_size) |
| 102 | max_size = f->real_file_size; |
| 103 | |
| 104 | max_blocks = max_size / (unsigned long long) td->o.min_bs[ddir]; |
| 105 | if (!max_blocks) |
| 106 | return 0; |
| 107 | |
| 108 | return max_blocks; |
| 109 | } |
| 110 | |
| 111 | /* |
| 112 | * Return the next free block in the map. |
| 113 | */ |
| 114 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, |
| 115 | enum fio_ddir ddir, unsigned long long *b) |
| 116 | { |
| 117 | unsigned long long min_bs = td->o.rw_min_bs; |
| 118 | int i; |
| 119 | |
| 120 | i = f->last_free_lookup; |
| 121 | *b = (i * BLOCKS_PER_MAP); |
| 122 | while ((*b) * min_bs < f->real_file_size) { |
| 123 | if (f->file_map[i] != (unsigned int) -1) { |
| 124 | *b += ffz(f->file_map[i]); |
| 125 | if (*b > last_block(td, f, ddir)) |
| 126 | break; |
| 127 | f->last_free_lookup = i; |
| 128 | return 0; |
| 129 | } |
| 130 | |
| 131 | *b += BLOCKS_PER_MAP; |
| 132 | i++; |
| 133 | } |
| 134 | |
| 135 | dprint(FD_IO, "failed finding a free block\n"); |
| 136 | return 1; |
| 137 | } |
| 138 | |
| 139 | static int get_next_rand_offset(struct thread_data *td, struct fio_file *f, |
| 140 | enum fio_ddir ddir, unsigned long long *b) |
| 141 | { |
| 142 | unsigned long long r; |
| 143 | int loops = 5; |
| 144 | |
| 145 | do { |
| 146 | r = os_random_long(&td->random_state); |
| 147 | dprint(FD_RANDOM, "off rand %llu\n", r); |
| 148 | *b = (last_block(td, f, ddir) - 1) |
| 149 | * (r / ((unsigned long long) RAND_MAX + 1.0)); |
| 150 | |
| 151 | /* |
| 152 | * if we are not maintaining a random map, we are done. |
| 153 | */ |
| 154 | if (!file_randommap(td, f)) |
| 155 | return 0; |
| 156 | |
| 157 | /* |
| 158 | * calculate map offset and check if it's free |
| 159 | */ |
| 160 | if (random_map_free(f, *b)) |
| 161 | return 0; |
| 162 | |
| 163 | dprint(FD_RANDOM, "get_next_rand_offset: offset %llu busy\n", |
| 164 | *b); |
| 165 | } while (--loops); |
| 166 | |
| 167 | /* |
| 168 | * we get here, if we didn't suceed in looking up a block. generate |
| 169 | * a random start offset into the filemap, and find the first free |
| 170 | * block from there. |
| 171 | */ |
| 172 | loops = 10; |
| 173 | do { |
| 174 | f->last_free_lookup = (f->num_maps - 1) * (r / (RAND_MAX+1.0)); |
| 175 | if (!get_next_free_block(td, f, ddir, b)) |
| 176 | return 0; |
| 177 | |
| 178 | r = os_random_long(&td->random_state); |
| 179 | } while (--loops); |
| 180 | |
| 181 | /* |
| 182 | * that didn't work either, try exhaustive search from the start |
| 183 | */ |
| 184 | f->last_free_lookup = 0; |
| 185 | return get_next_free_block(td, f, ddir, b); |
| 186 | } |
| 187 | |
| 188 | /* |
| 189 | * For random io, generate a random new block and see if it's used. Repeat |
| 190 | * until we find a free one. For sequential io, just return the end of |
| 191 | * the last io issued. |
| 192 | */ |
| 193 | static int get_next_offset(struct thread_data *td, struct io_u *io_u) |
| 194 | { |
| 195 | struct fio_file *f = io_u->file; |
| 196 | unsigned long long b; |
| 197 | enum fio_ddir ddir = io_u->ddir; |
| 198 | |
| 199 | if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) { |
| 200 | td->ddir_nr = td->o.ddir_nr; |
| 201 | |
| 202 | if (get_next_rand_offset(td, f, ddir, &b)) |
| 203 | return 1; |
| 204 | } else { |
| 205 | if (f->last_pos >= f->real_file_size) { |
| 206 | if (!td_random(td) || |
| 207 | get_next_rand_offset(td, f, ddir, &b)) |
| 208 | return 1; |
| 209 | } else |
| 210 | b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir]; |
| 211 | } |
| 212 | |
| 213 | io_u->offset = b * td->o.min_bs[ddir]; |
| 214 | if (io_u->offset >= f->io_size) { |
| 215 | dprint(FD_IO, "get_next_offset: offset %llu >= io_size %llu\n", |
| 216 | io_u->offset, f->io_size); |
| 217 | return 1; |
| 218 | } |
| 219 | |
| 220 | io_u->offset += f->file_offset; |
| 221 | if (io_u->offset >= f->real_file_size) { |
| 222 | dprint(FD_IO, "get_next_offset: offset %llu >= size %llu\n", |
| 223 | io_u->offset, f->real_file_size); |
| 224 | return 1; |
| 225 | } |
| 226 | |
| 227 | return 0; |
| 228 | } |
| 229 | |
| 230 | static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u) |
| 231 | { |
| 232 | const int ddir = io_u->ddir; |
| 233 | unsigned int buflen = buflen; /* silence dumb gcc warning */ |
| 234 | long r; |
| 235 | |
| 236 | if (td->o.min_bs[ddir] == td->o.max_bs[ddir]) |
| 237 | buflen = td->o.min_bs[ddir]; |
| 238 | else { |
| 239 | r = os_random_long(&td->bsrange_state); |
| 240 | if (!td->o.bssplit_nr) { |
| 241 | buflen = (unsigned int) |
| 242 | (1 + (double) (td->o.max_bs[ddir] - 1) |
| 243 | * r / (RAND_MAX + 1.0)); |
| 244 | } else { |
| 245 | long perc = 0; |
| 246 | unsigned int i; |
| 247 | |
| 248 | for (i = 0; i < td->o.bssplit_nr; i++) { |
| 249 | struct bssplit *bsp = &td->o.bssplit[i]; |
| 250 | |
| 251 | buflen = bsp->bs; |
| 252 | perc += bsp->perc; |
| 253 | if (r <= ((LONG_MAX / 100L) * perc)) |
| 254 | break; |
| 255 | } |
| 256 | } |
| 257 | if (!td->o.bs_unaligned) { |
| 258 | buflen = (buflen + td->o.min_bs[ddir] - 1) |
| 259 | & ~(td->o.min_bs[ddir] - 1); |
| 260 | } |
| 261 | } |
| 262 | |
| 263 | if (io_u->offset + buflen > io_u->file->real_file_size) { |
| 264 | dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen, |
| 265 | td->o.min_bs[ddir], ddir); |
| 266 | buflen = td->o.min_bs[ddir]; |
| 267 | } |
| 268 | |
| 269 | return buflen; |
| 270 | } |
| 271 | |
| 272 | static void set_rwmix_bytes(struct thread_data *td) |
| 273 | { |
| 274 | unsigned int diff; |
| 275 | |
| 276 | /* |
| 277 | * we do time or byte based switch. this is needed because |
| 278 | * buffered writes may issue a lot quicker than they complete, |
| 279 | * whereas reads do not. |
| 280 | */ |
| 281 | diff = td->o.rwmix[td->rwmix_ddir ^ 1]; |
| 282 | td->rwmix_issues = (td->io_issues[td->rwmix_ddir] * diff) / 100; |
| 283 | } |
| 284 | |
| 285 | static inline enum fio_ddir get_rand_ddir(struct thread_data *td) |
| 286 | { |
| 287 | unsigned int v; |
| 288 | long r; |
| 289 | |
| 290 | r = os_random_long(&td->rwmix_state); |
| 291 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); |
| 292 | if (v <= td->o.rwmix[DDIR_READ]) |
| 293 | return DDIR_READ; |
| 294 | |
| 295 | return DDIR_WRITE; |
| 296 | } |
| 297 | |
| 298 | /* |
| 299 | * Return the data direction for the next io_u. If the job is a |
| 300 | * mixed read/write workload, check the rwmix cycle and switch if |
| 301 | * necessary. |
| 302 | */ |
| 303 | static enum fio_ddir get_rw_ddir(struct thread_data *td) |
| 304 | { |
| 305 | if (td_rw(td)) { |
| 306 | /* |
| 307 | * Check if it's time to seed a new data direction. |
| 308 | */ |
| 309 | if (td->io_issues[td->rwmix_ddir] >= td->rwmix_issues) { |
| 310 | unsigned long long max_bytes; |
| 311 | enum fio_ddir ddir; |
| 312 | |
| 313 | /* |
| 314 | * Put a top limit on how many bytes we do for |
| 315 | * one data direction, to avoid overflowing the |
| 316 | * ranges too much |
| 317 | */ |
| 318 | ddir = get_rand_ddir(td); |
| 319 | max_bytes = td->this_io_bytes[ddir]; |
| 320 | if (max_bytes >= |
| 321 | (td->o.size * td->o.rwmix[ddir] / 100)) { |
| 322 | if (!td->rw_end_set[ddir]) |
| 323 | td->rw_end_set[ddir] = 1; |
| 324 | |
| 325 | ddir ^= 1; |
| 326 | } |
| 327 | |
| 328 | if (ddir != td->rwmix_ddir) |
| 329 | set_rwmix_bytes(td); |
| 330 | |
| 331 | td->rwmix_ddir = ddir; |
| 332 | } |
| 333 | return td->rwmix_ddir; |
| 334 | } else if (td_read(td)) |
| 335 | return DDIR_READ; |
| 336 | else |
| 337 | return DDIR_WRITE; |
| 338 | } |
| 339 | |
| 340 | static void put_file_log(struct thread_data *td, struct fio_file *f) |
| 341 | { |
| 342 | int ret = put_file(td, f); |
| 343 | |
| 344 | if (ret) |
| 345 | td_verror(td, ret, "file close"); |
| 346 | } |
| 347 | |
| 348 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
| 349 | { |
| 350 | assert((io_u->flags & IO_U_F_FREE) == 0); |
| 351 | io_u->flags |= IO_U_F_FREE; |
| 352 | |
| 353 | if (io_u->file) |
| 354 | put_file_log(td, io_u->file); |
| 355 | |
| 356 | io_u->file = NULL; |
| 357 | list_del(&io_u->list); |
| 358 | list_add(&io_u->list, &td->io_u_freelist); |
| 359 | td->cur_depth--; |
| 360 | } |
| 361 | |
| 362 | void requeue_io_u(struct thread_data *td, struct io_u **io_u) |
| 363 | { |
| 364 | struct io_u *__io_u = *io_u; |
| 365 | |
| 366 | dprint(FD_IO, "requeue %p\n", __io_u); |
| 367 | |
| 368 | __io_u->flags |= IO_U_F_FREE; |
| 369 | if ((__io_u->flags & IO_U_F_FLIGHT) && (__io_u->ddir != DDIR_SYNC)) |
| 370 | td->io_issues[__io_u->ddir]--; |
| 371 | |
| 372 | __io_u->flags &= ~IO_U_F_FLIGHT; |
| 373 | |
| 374 | list_del(&__io_u->list); |
| 375 | list_add_tail(&__io_u->list, &td->io_u_requeues); |
| 376 | td->cur_depth--; |
| 377 | *io_u = NULL; |
| 378 | } |
| 379 | |
| 380 | static int fill_io_u(struct thread_data *td, struct io_u *io_u) |
| 381 | { |
| 382 | if (td->io_ops->flags & FIO_NOIO) |
| 383 | goto out; |
| 384 | |
| 385 | /* |
| 386 | * see if it's time to sync |
| 387 | */ |
| 388 | if (td->o.fsync_blocks && |
| 389 | !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) && |
| 390 | td->io_issues[DDIR_WRITE] && should_fsync(td)) { |
| 391 | io_u->ddir = DDIR_SYNC; |
| 392 | goto out; |
| 393 | } |
| 394 | |
| 395 | io_u->ddir = get_rw_ddir(td); |
| 396 | |
| 397 | /* |
| 398 | * See if it's time to switch to a new zone |
| 399 | */ |
| 400 | if (td->zone_bytes >= td->o.zone_size) { |
| 401 | td->zone_bytes = 0; |
| 402 | io_u->file->last_pos += td->o.zone_skip; |
| 403 | td->io_skip_bytes += td->o.zone_skip; |
| 404 | } |
| 405 | |
| 406 | /* |
| 407 | * No log, let the seq/rand engine retrieve the next buflen and |
| 408 | * position. |
| 409 | */ |
| 410 | if (get_next_offset(td, io_u)) { |
| 411 | dprint(FD_IO, "io_u %p, failed getting offset\n", io_u); |
| 412 | return 1; |
| 413 | } |
| 414 | |
| 415 | io_u->buflen = get_next_buflen(td, io_u); |
| 416 | if (!io_u->buflen) { |
| 417 | dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u); |
| 418 | return 1; |
| 419 | } |
| 420 | |
| 421 | if (io_u->offset + io_u->buflen > io_u->file->real_file_size) { |
| 422 | dprint(FD_IO, "io_u %p, offset too large\n", io_u); |
| 423 | dprint(FD_IO, " off=%llu/%lu > %llu\n", io_u->offset, |
| 424 | io_u->buflen, io_u->file->real_file_size); |
| 425 | return 1; |
| 426 | } |
| 427 | |
| 428 | /* |
| 429 | * mark entry before potentially trimming io_u |
| 430 | */ |
| 431 | if (td_random(td) && file_randommap(td, io_u->file)) |
| 432 | mark_random_map(td, io_u); |
| 433 | |
| 434 | /* |
| 435 | * If using a write iolog, store this entry. |
| 436 | */ |
| 437 | out: |
| 438 | dprint_io_u(io_u, "fill_io_u"); |
| 439 | td->zone_bytes += io_u->buflen; |
| 440 | log_io_u(td, io_u); |
| 441 | return 0; |
| 442 | } |
| 443 | |
| 444 | static void __io_u_mark_map(unsigned int *map, unsigned int nr) |
| 445 | { |
| 446 | int index = 0; |
| 447 | |
| 448 | switch (nr) { |
| 449 | default: |
| 450 | index = 6; |
| 451 | break; |
| 452 | case 33 ... 64: |
| 453 | index = 5; |
| 454 | break; |
| 455 | case 17 ... 32: |
| 456 | index = 4; |
| 457 | break; |
| 458 | case 9 ... 16: |
| 459 | index = 3; |
| 460 | break; |
| 461 | case 5 ... 8: |
| 462 | index = 2; |
| 463 | break; |
| 464 | case 1 ... 4: |
| 465 | index = 1; |
| 466 | case 0: |
| 467 | break; |
| 468 | } |
| 469 | |
| 470 | map[index]++; |
| 471 | } |
| 472 | |
| 473 | void io_u_mark_submit(struct thread_data *td, unsigned int nr) |
| 474 | { |
| 475 | __io_u_mark_map(td->ts.io_u_submit, nr); |
| 476 | td->ts.total_submit++; |
| 477 | } |
| 478 | |
| 479 | void io_u_mark_complete(struct thread_data *td, unsigned int nr) |
| 480 | { |
| 481 | __io_u_mark_map(td->ts.io_u_complete, nr); |
| 482 | td->ts.total_complete++; |
| 483 | } |
| 484 | |
| 485 | void io_u_mark_depth(struct thread_data *td, unsigned int nr) |
| 486 | { |
| 487 | int index = 0; |
| 488 | |
| 489 | switch (td->cur_depth) { |
| 490 | default: |
| 491 | index = 6; |
| 492 | break; |
| 493 | case 32 ... 63: |
| 494 | index = 5; |
| 495 | break; |
| 496 | case 16 ... 31: |
| 497 | index = 4; |
| 498 | break; |
| 499 | case 8 ... 15: |
| 500 | index = 3; |
| 501 | break; |
| 502 | case 4 ... 7: |
| 503 | index = 2; |
| 504 | break; |
| 505 | case 2 ... 3: |
| 506 | index = 1; |
| 507 | case 1: |
| 508 | break; |
| 509 | } |
| 510 | |
| 511 | td->ts.io_u_map[index] += nr; |
| 512 | } |
| 513 | |
| 514 | static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec) |
| 515 | { |
| 516 | int index = 0; |
| 517 | |
| 518 | assert(usec < 1000); |
| 519 | |
| 520 | switch (usec) { |
| 521 | case 750 ... 999: |
| 522 | index = 9; |
| 523 | break; |
| 524 | case 500 ... 749: |
| 525 | index = 8; |
| 526 | break; |
| 527 | case 250 ... 499: |
| 528 | index = 7; |
| 529 | break; |
| 530 | case 100 ... 249: |
| 531 | index = 6; |
| 532 | break; |
| 533 | case 50 ... 99: |
| 534 | index = 5; |
| 535 | break; |
| 536 | case 20 ... 49: |
| 537 | index = 4; |
| 538 | break; |
| 539 | case 10 ... 19: |
| 540 | index = 3; |
| 541 | break; |
| 542 | case 4 ... 9: |
| 543 | index = 2; |
| 544 | break; |
| 545 | case 2 ... 3: |
| 546 | index = 1; |
| 547 | case 0 ... 1: |
| 548 | break; |
| 549 | } |
| 550 | |
| 551 | assert(index < FIO_IO_U_LAT_U_NR); |
| 552 | td->ts.io_u_lat_u[index]++; |
| 553 | } |
| 554 | |
| 555 | static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec) |
| 556 | { |
| 557 | int index = 0; |
| 558 | |
| 559 | switch (msec) { |
| 560 | default: |
| 561 | index = 11; |
| 562 | break; |
| 563 | case 1000 ... 1999: |
| 564 | index = 10; |
| 565 | break; |
| 566 | case 750 ... 999: |
| 567 | index = 9; |
| 568 | break; |
| 569 | case 500 ... 749: |
| 570 | index = 8; |
| 571 | break; |
| 572 | case 250 ... 499: |
| 573 | index = 7; |
| 574 | break; |
| 575 | case 100 ... 249: |
| 576 | index = 6; |
| 577 | break; |
| 578 | case 50 ... 99: |
| 579 | index = 5; |
| 580 | break; |
| 581 | case 20 ... 49: |
| 582 | index = 4; |
| 583 | break; |
| 584 | case 10 ... 19: |
| 585 | index = 3; |
| 586 | break; |
| 587 | case 4 ... 9: |
| 588 | index = 2; |
| 589 | break; |
| 590 | case 2 ... 3: |
| 591 | index = 1; |
| 592 | case 0 ... 1: |
| 593 | break; |
| 594 | } |
| 595 | |
| 596 | assert(index < FIO_IO_U_LAT_M_NR); |
| 597 | td->ts.io_u_lat_m[index]++; |
| 598 | } |
| 599 | |
| 600 | static void io_u_mark_latency(struct thread_data *td, unsigned long usec) |
| 601 | { |
| 602 | if (usec < 1000) |
| 603 | io_u_mark_lat_usec(td, usec); |
| 604 | else |
| 605 | io_u_mark_lat_msec(td, usec / 1000); |
| 606 | } |
| 607 | |
| 608 | /* |
| 609 | * Get next file to service by choosing one at random |
| 610 | */ |
| 611 | static struct fio_file *get_next_file_rand(struct thread_data *td, int goodf, |
| 612 | int badf) |
| 613 | { |
| 614 | struct fio_file *f; |
| 615 | int fno; |
| 616 | |
| 617 | do { |
| 618 | long r = os_random_long(&td->next_file_state); |
| 619 | |
| 620 | fno = (unsigned int) ((double) td->o.nr_files |
| 621 | * (r / (RAND_MAX + 1.0))); |
| 622 | f = td->files[fno]; |
| 623 | if (f->flags & FIO_FILE_DONE) |
| 624 | continue; |
| 625 | |
| 626 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) { |
| 627 | dprint(FD_FILE, "get_next_file_rand: %p\n", f); |
| 628 | return f; |
| 629 | } |
| 630 | } while (1); |
| 631 | } |
| 632 | |
| 633 | /* |
| 634 | * Get next file to service by doing round robin between all available ones |
| 635 | */ |
| 636 | static struct fio_file *get_next_file_rr(struct thread_data *td, int goodf, |
| 637 | int badf) |
| 638 | { |
| 639 | unsigned int old_next_file = td->next_file; |
| 640 | struct fio_file *f; |
| 641 | |
| 642 | do { |
| 643 | f = td->files[td->next_file]; |
| 644 | |
| 645 | td->next_file++; |
| 646 | if (td->next_file >= td->o.nr_files) |
| 647 | td->next_file = 0; |
| 648 | |
| 649 | if (f->flags & FIO_FILE_DONE) { |
| 650 | f = NULL; |
| 651 | continue; |
| 652 | } |
| 653 | |
| 654 | if ((!goodf || (f->flags & goodf)) && !(f->flags & badf)) |
| 655 | break; |
| 656 | |
| 657 | f = NULL; |
| 658 | } while (td->next_file != old_next_file); |
| 659 | |
| 660 | dprint(FD_FILE, "get_next_file_rr: %p\n", f); |
| 661 | return f; |
| 662 | } |
| 663 | |
| 664 | static struct fio_file *get_next_file(struct thread_data *td) |
| 665 | { |
| 666 | struct fio_file *f; |
| 667 | |
| 668 | assert(td->o.nr_files <= td->files_index); |
| 669 | |
| 670 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) { |
| 671 | dprint(FD_FILE, "get_next_file: nr_open=%d, nr_done=%d," |
| 672 | " nr_files=%d\n", td->nr_open_files, |
| 673 | td->nr_done_files, |
| 674 | td->o.nr_files); |
| 675 | return NULL; |
| 676 | } |
| 677 | |
| 678 | f = td->file_service_file; |
| 679 | if (f && (f->flags & FIO_FILE_OPEN) && td->file_service_left--) |
| 680 | goto out; |
| 681 | |
| 682 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
| 683 | f = get_next_file_rr(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
| 684 | else |
| 685 | f = get_next_file_rand(td, FIO_FILE_OPEN, FIO_FILE_CLOSING); |
| 686 | |
| 687 | td->file_service_file = f; |
| 688 | td->file_service_left = td->file_service_nr - 1; |
| 689 | out: |
| 690 | dprint(FD_FILE, "get_next_file: %p\n", f); |
| 691 | return f; |
| 692 | } |
| 693 | |
| 694 | static struct fio_file *find_next_new_file(struct thread_data *td) |
| 695 | { |
| 696 | struct fio_file *f; |
| 697 | |
| 698 | if (!td->nr_open_files || td->nr_done_files >= td->o.nr_files) |
| 699 | return NULL; |
| 700 | |
| 701 | if (td->o.file_service_type == FIO_FSERVICE_RR) |
| 702 | f = get_next_file_rr(td, 0, FIO_FILE_OPEN); |
| 703 | else |
| 704 | f = get_next_file_rand(td, 0, FIO_FILE_OPEN); |
| 705 | |
| 706 | return f; |
| 707 | } |
| 708 | |
| 709 | static int set_io_u_file(struct thread_data *td, struct io_u *io_u) |
| 710 | { |
| 711 | struct fio_file *f; |
| 712 | |
| 713 | do { |
| 714 | f = get_next_file(td); |
| 715 | if (!f) |
| 716 | return 1; |
| 717 | |
| 718 | set_file: |
| 719 | io_u->file = f; |
| 720 | get_file(f); |
| 721 | |
| 722 | if (!fill_io_u(td, io_u)) |
| 723 | break; |
| 724 | |
| 725 | /* |
| 726 | * optimization to prevent close/open of the same file. This |
| 727 | * way we preserve queueing etc. |
| 728 | */ |
| 729 | if (td->o.nr_files == 1 && td->o.time_based) { |
| 730 | put_file_log(td, f); |
| 731 | fio_file_reset(f); |
| 732 | goto set_file; |
| 733 | } |
| 734 | |
| 735 | /* |
| 736 | * td_io_close() does a put_file() as well, so no need to |
| 737 | * do that here. |
| 738 | */ |
| 739 | io_u->file = NULL; |
| 740 | td_io_close_file(td, f); |
| 741 | f->flags |= FIO_FILE_DONE; |
| 742 | td->nr_done_files++; |
| 743 | |
| 744 | /* |
| 745 | * probably not the right place to do this, but see |
| 746 | * if we need to open a new file |
| 747 | */ |
| 748 | if (td->nr_open_files < td->o.open_files && |
| 749 | td->o.open_files != td->o.nr_files) { |
| 750 | f = find_next_new_file(td); |
| 751 | |
| 752 | if (!f || td_io_open_file(td, f)) |
| 753 | return 1; |
| 754 | |
| 755 | goto set_file; |
| 756 | } |
| 757 | } while (1); |
| 758 | |
| 759 | return 0; |
| 760 | } |
| 761 | |
| 762 | |
| 763 | struct io_u *__get_io_u(struct thread_data *td) |
| 764 | { |
| 765 | struct io_u *io_u = NULL; |
| 766 | |
| 767 | if (!list_empty(&td->io_u_requeues)) |
| 768 | io_u = list_entry(td->io_u_requeues.next, struct io_u, list); |
| 769 | else if (!queue_full(td)) { |
| 770 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); |
| 771 | |
| 772 | io_u->buflen = 0; |
| 773 | io_u->resid = 0; |
| 774 | io_u->file = NULL; |
| 775 | io_u->end_io = NULL; |
| 776 | } |
| 777 | |
| 778 | if (io_u) { |
| 779 | assert(io_u->flags & IO_U_F_FREE); |
| 780 | io_u->flags &= ~IO_U_F_FREE; |
| 781 | |
| 782 | io_u->error = 0; |
| 783 | list_del(&io_u->list); |
| 784 | list_add(&io_u->list, &td->io_u_busylist); |
| 785 | td->cur_depth++; |
| 786 | } |
| 787 | |
| 788 | return io_u; |
| 789 | } |
| 790 | |
| 791 | /* |
| 792 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, |
| 793 | * etc. The returned io_u is fully ready to be prepped and submitted. |
| 794 | */ |
| 795 | struct io_u *get_io_u(struct thread_data *td) |
| 796 | { |
| 797 | struct fio_file *f; |
| 798 | struct io_u *io_u; |
| 799 | |
| 800 | io_u = __get_io_u(td); |
| 801 | if (!io_u) { |
| 802 | dprint(FD_IO, "__get_io_u failed\n"); |
| 803 | return NULL; |
| 804 | } |
| 805 | |
| 806 | /* |
| 807 | * from a requeue, io_u already setup |
| 808 | */ |
| 809 | if (io_u->file) |
| 810 | goto out; |
| 811 | |
| 812 | /* |
| 813 | * If using an iolog, grab next piece if any available. |
| 814 | */ |
| 815 | if (td->o.read_iolog_file) { |
| 816 | if (read_iolog_get(td, io_u)) |
| 817 | goto err_put; |
| 818 | } else if (set_io_u_file(td, io_u)) { |
| 819 | dprint(FD_IO, "io_u %p, setting file failed\n", io_u); |
| 820 | goto err_put; |
| 821 | } |
| 822 | |
| 823 | f = io_u->file; |
| 824 | assert(f->flags & FIO_FILE_OPEN); |
| 825 | |
| 826 | if (io_u->ddir != DDIR_SYNC) { |
| 827 | if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) { |
| 828 | dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u); |
| 829 | goto err_put; |
| 830 | } |
| 831 | |
| 832 | f->last_pos = io_u->offset + io_u->buflen; |
| 833 | |
| 834 | if (td->o.verify != VERIFY_NONE) |
| 835 | populate_verify_io_u(td, io_u); |
| 836 | else if (td->o.refill_buffers && io_u->ddir == DDIR_WRITE) |
| 837 | io_u_fill_buffer(td, io_u, io_u->xfer_buflen); |
| 838 | } |
| 839 | |
| 840 | /* |
| 841 | * Set io data pointers. |
| 842 | */ |
| 843 | io_u->endpos = io_u->offset + io_u->buflen; |
| 844 | io_u->xfer_buf = io_u->buf; |
| 845 | io_u->xfer_buflen = io_u->buflen; |
| 846 | |
| 847 | out: |
| 848 | if (!td_io_prep(td, io_u)) { |
| 849 | fio_gettime(&io_u->start_time, NULL); |
| 850 | return io_u; |
| 851 | } |
| 852 | err_put: |
| 853 | dprint(FD_IO, "get_io_u failed\n"); |
| 854 | put_io_u(td, io_u); |
| 855 | return NULL; |
| 856 | } |
| 857 | |
| 858 | void io_u_log_error(struct thread_data *td, struct io_u *io_u) |
| 859 | { |
| 860 | const char *msg[] = { "read", "write", "sync" }; |
| 861 | |
| 862 | log_err("fio: io_u error"); |
| 863 | |
| 864 | if (io_u->file) |
| 865 | log_err(" on file %s", io_u->file->file_name); |
| 866 | |
| 867 | log_err(": %s\n", strerror(io_u->error)); |
| 868 | |
| 869 | log_err(" %s offset=%llu, buflen=%lu\n", msg[io_u->ddir], |
| 870 | io_u->offset, io_u->xfer_buflen); |
| 871 | |
| 872 | if (!td->error) |
| 873 | td_verror(td, io_u->error, "io_u error"); |
| 874 | } |
| 875 | |
| 876 | static void io_completed(struct thread_data *td, struct io_u *io_u, |
| 877 | struct io_completion_data *icd) |
| 878 | { |
| 879 | unsigned long usec; |
| 880 | |
| 881 | dprint_io_u(io_u, "io complete"); |
| 882 | |
| 883 | assert(io_u->flags & IO_U_F_FLIGHT); |
| 884 | io_u->flags &= ~IO_U_F_FLIGHT; |
| 885 | |
| 886 | if (io_u->ddir == DDIR_SYNC) { |
| 887 | td->last_was_sync = 1; |
| 888 | return; |
| 889 | } |
| 890 | |
| 891 | td->last_was_sync = 0; |
| 892 | |
| 893 | if (!io_u->error) { |
| 894 | unsigned int bytes = io_u->buflen - io_u->resid; |
| 895 | const enum fio_ddir idx = io_u->ddir; |
| 896 | int ret; |
| 897 | |
| 898 | td->io_blocks[idx]++; |
| 899 | td->io_bytes[idx] += bytes; |
| 900 | td->this_io_bytes[idx] += bytes; |
| 901 | |
| 902 | usec = utime_since(&io_u->issue_time, &icd->time); |
| 903 | |
| 904 | add_clat_sample(td, idx, usec); |
| 905 | add_bw_sample(td, idx, &icd->time); |
| 906 | io_u_mark_latency(td, usec); |
| 907 | |
| 908 | if (td_write(td) && idx == DDIR_WRITE && |
| 909 | td->o.do_verify && |
| 910 | td->o.verify != VERIFY_NONE) |
| 911 | log_io_piece(td, io_u); |
| 912 | |
| 913 | icd->bytes_done[idx] += bytes; |
| 914 | |
| 915 | if (io_u->end_io) { |
| 916 | ret = io_u->end_io(td, io_u); |
| 917 | if (ret && !icd->error) |
| 918 | icd->error = ret; |
| 919 | } |
| 920 | } else { |
| 921 | icd->error = io_u->error; |
| 922 | io_u_log_error(td, io_u); |
| 923 | } |
| 924 | } |
| 925 | |
| 926 | static void init_icd(struct io_completion_data *icd, int nr) |
| 927 | { |
| 928 | fio_gettime(&icd->time, NULL); |
| 929 | |
| 930 | icd->nr = nr; |
| 931 | |
| 932 | icd->error = 0; |
| 933 | icd->bytes_done[0] = icd->bytes_done[1] = 0; |
| 934 | } |
| 935 | |
| 936 | static void ios_completed(struct thread_data *td, |
| 937 | struct io_completion_data *icd) |
| 938 | { |
| 939 | struct io_u *io_u; |
| 940 | int i; |
| 941 | |
| 942 | for (i = 0; i < icd->nr; i++) { |
| 943 | io_u = td->io_ops->event(td, i); |
| 944 | |
| 945 | io_completed(td, io_u, icd); |
| 946 | put_io_u(td, io_u); |
| 947 | } |
| 948 | } |
| 949 | |
| 950 | /* |
| 951 | * Complete a single io_u for the sync engines. |
| 952 | */ |
| 953 | long io_u_sync_complete(struct thread_data *td, struct io_u *io_u) |
| 954 | { |
| 955 | struct io_completion_data icd; |
| 956 | |
| 957 | init_icd(&icd, 1); |
| 958 | io_completed(td, io_u, &icd); |
| 959 | put_io_u(td, io_u); |
| 960 | |
| 961 | if (!icd.error) |
| 962 | return icd.bytes_done[0] + icd.bytes_done[1]; |
| 963 | |
| 964 | td_verror(td, icd.error, "io_u_sync_complete"); |
| 965 | return -1; |
| 966 | } |
| 967 | |
| 968 | /* |
| 969 | * Called to complete min_events number of io for the async engines. |
| 970 | */ |
| 971 | long io_u_queued_complete(struct thread_data *td, int min_events) |
| 972 | { |
| 973 | struct io_completion_data icd; |
| 974 | struct timespec *tvp = NULL; |
| 975 | int ret; |
| 976 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, }; |
| 977 | |
| 978 | dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_events); |
| 979 | |
| 980 | if (!min_events) |
| 981 | tvp = &ts; |
| 982 | |
| 983 | ret = td_io_getevents(td, min_events, td->cur_depth, tvp); |
| 984 | if (ret < 0) { |
| 985 | td_verror(td, -ret, "td_io_getevents"); |
| 986 | return ret; |
| 987 | } else if (!ret) |
| 988 | return ret; |
| 989 | |
| 990 | init_icd(&icd, ret); |
| 991 | ios_completed(td, &icd); |
| 992 | if (!icd.error) |
| 993 | return icd.bytes_done[0] + icd.bytes_done[1]; |
| 994 | |
| 995 | td_verror(td, icd.error, "io_u_queued_complete"); |
| 996 | return -1; |
| 997 | } |
| 998 | |
| 999 | /* |
| 1000 | * Call when io_u is really queued, to update the submission latency. |
| 1001 | */ |
| 1002 | void io_u_queued(struct thread_data *td, struct io_u *io_u) |
| 1003 | { |
| 1004 | unsigned long slat_time; |
| 1005 | |
| 1006 | slat_time = utime_since(&io_u->start_time, &io_u->issue_time); |
| 1007 | add_slat_sample(td, io_u->ddir, slat_time); |
| 1008 | } |
| 1009 | |
| 1010 | /* |
| 1011 | * "randomly" fill the buffer contents |
| 1012 | */ |
| 1013 | void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u, |
| 1014 | unsigned int max_bs) |
| 1015 | { |
| 1016 | long *ptr = io_u->buf; |
| 1017 | |
| 1018 | if (!td->o.zero_buffers) { |
| 1019 | while ((void *) ptr - io_u->buf < max_bs) { |
| 1020 | *ptr = rand() * GOLDEN_RATIO_PRIME; |
| 1021 | ptr++; |
| 1022 | } |
| 1023 | } else |
| 1024 | memset(ptr, 0, max_bs); |
| 1025 | } |
| 1026 | |
| 1027 | #ifdef FIO_USE_TIMEOUT |
| 1028 | void io_u_set_timeout(struct thread_data *td) |
| 1029 | { |
| 1030 | assert(td->cur_depth); |
| 1031 | |
| 1032 | td->timer.it_interval.tv_sec = 0; |
| 1033 | td->timer.it_interval.tv_usec = 0; |
| 1034 | td->timer.it_value.tv_sec = IO_U_TIMEOUT + IO_U_TIMEOUT_INC; |
| 1035 | td->timer.it_value.tv_usec = 0; |
| 1036 | setitimer(ITIMER_REAL, &td->timer, NULL); |
| 1037 | fio_gettime(&td->timeout_end, NULL); |
| 1038 | } |
| 1039 | |
| 1040 | static void io_u_dump(struct io_u *io_u) |
| 1041 | { |
| 1042 | unsigned long t_start = mtime_since_now(&io_u->start_time); |
| 1043 | unsigned long t_issue = mtime_since_now(&io_u->issue_time); |
| 1044 | |
| 1045 | log_err("io_u=%p, t_start=%lu, t_issue=%lu\n", io_u, t_start, t_issue); |
| 1046 | log_err(" buf=%p/%p, len=%lu/%lu, offset=%llu\n", io_u->buf, |
| 1047 | io_u->xfer_buf, io_u->buflen, |
| 1048 | io_u->xfer_buflen, |
| 1049 | io_u->offset); |
| 1050 | log_err(" ddir=%d, fname=%s\n", io_u->ddir, io_u->file->file_name); |
| 1051 | } |
| 1052 | #else |
| 1053 | void io_u_set_timeout(struct thread_data fio_unused *td) |
| 1054 | { |
| 1055 | } |
| 1056 | #endif |
| 1057 | |
| 1058 | #ifdef FIO_USE_TIMEOUT |
| 1059 | static void io_u_timeout_handler(int fio_unused sig) |
| 1060 | { |
| 1061 | struct thread_data *td, *__td; |
| 1062 | pid_t pid = getpid(); |
| 1063 | struct list_head *entry; |
| 1064 | struct io_u *io_u; |
| 1065 | int i; |
| 1066 | |
| 1067 | log_err("fio: io_u timeout\n"); |
| 1068 | |
| 1069 | /* |
| 1070 | * TLS would be nice... |
| 1071 | */ |
| 1072 | td = NULL; |
| 1073 | for_each_td(__td, i) { |
| 1074 | if (__td->pid == pid) { |
| 1075 | td = __td; |
| 1076 | break; |
| 1077 | } |
| 1078 | } |
| 1079 | |
| 1080 | if (!td) { |
| 1081 | log_err("fio: io_u timeout, can't find job\n"); |
| 1082 | exit(1); |
| 1083 | } |
| 1084 | |
| 1085 | if (!td->cur_depth) { |
| 1086 | log_err("fio: timeout without pending work?\n"); |
| 1087 | return; |
| 1088 | } |
| 1089 | |
| 1090 | log_err("fio: io_u timeout: job=%s, pid=%d\n", td->o.name, td->pid); |
| 1091 | |
| 1092 | list_for_each(entry, &td->io_u_busylist) { |
| 1093 | io_u = list_entry(entry, struct io_u, list); |
| 1094 | |
| 1095 | io_u_dump(io_u); |
| 1096 | } |
| 1097 | |
| 1098 | td_verror(td, ETIMEDOUT, "io_u timeout"); |
| 1099 | exit(1); |
| 1100 | } |
| 1101 | #endif |
| 1102 | |
| 1103 | void io_u_init_timeout(void) |
| 1104 | { |
| 1105 | #ifdef FIO_USE_TIMEOUT |
| 1106 | signal(SIGALRM, io_u_timeout_handler); |
| 1107 | #endif |
| 1108 | } |