| 1 | #include <unistd.h> |
| 2 | #include <fcntl.h> |
| 3 | #include <string.h> |
| 4 | #include <assert.h> |
| 5 | #include <dirent.h> |
| 6 | #include <libgen.h> |
| 7 | #include <sys/stat.h> |
| 8 | #include <sys/mman.h> |
| 9 | #include <sys/types.h> |
| 10 | |
| 11 | #include "fio.h" |
| 12 | #include "smalloc.h" |
| 13 | #include "filehash.h" |
| 14 | #include "options.h" |
| 15 | #include "os/os.h" |
| 16 | #include "hash.h" |
| 17 | #include "lib/axmap.h" |
| 18 | |
| 19 | #ifdef CONFIG_LINUX_FALLOCATE |
| 20 | #include <linux/falloc.h> |
| 21 | #endif |
| 22 | |
| 23 | static int root_warn; |
| 24 | |
| 25 | static FLIST_HEAD(filename_list); |
| 26 | |
| 27 | /* |
| 28 | * List entry for filename_list |
| 29 | */ |
| 30 | struct file_name { |
| 31 | struct flist_head list; |
| 32 | char *filename; |
| 33 | }; |
| 34 | |
| 35 | static inline void clear_error(struct thread_data *td) |
| 36 | { |
| 37 | td->error = 0; |
| 38 | td->verror[0] = '\0'; |
| 39 | } |
| 40 | |
| 41 | /* |
| 42 | * Leaves f->fd open on success, caller must close |
| 43 | */ |
| 44 | static int extend_file(struct thread_data *td, struct fio_file *f) |
| 45 | { |
| 46 | int r, new_layout = 0, unlink_file = 0, flags; |
| 47 | unsigned long long left; |
| 48 | unsigned int bs; |
| 49 | char *b = NULL; |
| 50 | |
| 51 | if (read_only) { |
| 52 | log_err("fio: refusing extend of file due to read-only\n"); |
| 53 | return 0; |
| 54 | } |
| 55 | |
| 56 | /* |
| 57 | * check if we need to lay the file out complete again. fio |
| 58 | * does that for operations involving reads, or for writes |
| 59 | * where overwrite is set |
| 60 | */ |
| 61 | if (td_read(td) || |
| 62 | (td_write(td) && td->o.overwrite && !td->o.file_append) || |
| 63 | (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND))) |
| 64 | new_layout = 1; |
| 65 | if (td_write(td) && !td->o.overwrite && !td->o.file_append) |
| 66 | unlink_file = 1; |
| 67 | |
| 68 | if (unlink_file || new_layout) { |
| 69 | int ret; |
| 70 | |
| 71 | dprint(FD_FILE, "layout unlink %s\n", f->file_name); |
| 72 | |
| 73 | ret = td_io_unlink_file(td, f); |
| 74 | if (ret != 0 && ret != ENOENT) { |
| 75 | td_verror(td, errno, "unlink"); |
| 76 | return 1; |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | flags = O_WRONLY; |
| 81 | if (td->o.allow_create) |
| 82 | flags |= O_CREAT; |
| 83 | if (new_layout) |
| 84 | flags |= O_TRUNC; |
| 85 | |
| 86 | #ifdef WIN32 |
| 87 | flags |= _O_BINARY; |
| 88 | #endif |
| 89 | |
| 90 | dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags); |
| 91 | f->fd = open(f->file_name, flags, 0644); |
| 92 | if (f->fd < 0) { |
| 93 | int err = errno; |
| 94 | |
| 95 | if (err == ENOENT && !td->o.allow_create) |
| 96 | log_err("fio: file creation disallowed by " |
| 97 | "allow_file_create=0\n"); |
| 98 | else |
| 99 | td_verror(td, err, "open"); |
| 100 | return 1; |
| 101 | } |
| 102 | |
| 103 | #ifdef CONFIG_POSIX_FALLOCATE |
| 104 | if (!td->o.fill_device) { |
| 105 | switch (td->o.fallocate_mode) { |
| 106 | case FIO_FALLOCATE_NONE: |
| 107 | break; |
| 108 | case FIO_FALLOCATE_POSIX: |
| 109 | dprint(FD_FILE, "posix_fallocate file %s size %llu\n", |
| 110 | f->file_name, |
| 111 | (unsigned long long) f->real_file_size); |
| 112 | |
| 113 | r = posix_fallocate(f->fd, 0, f->real_file_size); |
| 114 | if (r > 0) { |
| 115 | log_err("fio: posix_fallocate fails: %s\n", |
| 116 | strerror(r)); |
| 117 | } |
| 118 | break; |
| 119 | #ifdef CONFIG_LINUX_FALLOCATE |
| 120 | case FIO_FALLOCATE_KEEP_SIZE: |
| 121 | dprint(FD_FILE, |
| 122 | "fallocate(FALLOC_FL_KEEP_SIZE) " |
| 123 | "file %s size %llu\n", f->file_name, |
| 124 | (unsigned long long) f->real_file_size); |
| 125 | |
| 126 | r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, |
| 127 | f->real_file_size); |
| 128 | if (r != 0) |
| 129 | td_verror(td, errno, "fallocate"); |
| 130 | |
| 131 | break; |
| 132 | #endif /* CONFIG_LINUX_FALLOCATE */ |
| 133 | default: |
| 134 | log_err("fio: unknown fallocate mode: %d\n", |
| 135 | td->o.fallocate_mode); |
| 136 | assert(0); |
| 137 | } |
| 138 | } |
| 139 | #endif /* CONFIG_POSIX_FALLOCATE */ |
| 140 | |
| 141 | /* |
| 142 | * If our jobs don't require regular files initially, we're done. |
| 143 | */ |
| 144 | if (!new_layout) |
| 145 | goto done; |
| 146 | |
| 147 | /* |
| 148 | * The size will be -1ULL when fill_device is used, so don't truncate |
| 149 | * or fallocate this file, just write it |
| 150 | */ |
| 151 | if (!td->o.fill_device) { |
| 152 | dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name, |
| 153 | (unsigned long long) f->real_file_size); |
| 154 | if (ftruncate(f->fd, f->real_file_size) == -1) { |
| 155 | if (errno != EFBIG) { |
| 156 | td_verror(td, errno, "ftruncate"); |
| 157 | goto err; |
| 158 | } |
| 159 | } |
| 160 | } |
| 161 | |
| 162 | left = f->real_file_size; |
| 163 | bs = td->o.max_bs[DDIR_WRITE]; |
| 164 | if (bs > left) |
| 165 | bs = left; |
| 166 | |
| 167 | b = malloc(bs); |
| 168 | if (!b) { |
| 169 | td_verror(td, errno, "malloc"); |
| 170 | goto err; |
| 171 | } |
| 172 | |
| 173 | while (left && !td->terminate) { |
| 174 | if (bs > left) |
| 175 | bs = left; |
| 176 | |
| 177 | fill_io_buffer(td, b, bs, bs); |
| 178 | |
| 179 | r = write(f->fd, b, bs); |
| 180 | |
| 181 | if (r > 0) { |
| 182 | left -= r; |
| 183 | continue; |
| 184 | } else { |
| 185 | if (r < 0) { |
| 186 | int __e = errno; |
| 187 | |
| 188 | if (__e == ENOSPC) { |
| 189 | if (td->o.fill_device) |
| 190 | break; |
| 191 | log_info("fio: ENOSPC on laying out " |
| 192 | "file, stopping\n"); |
| 193 | break; |
| 194 | } |
| 195 | td_verror(td, errno, "write"); |
| 196 | } else |
| 197 | td_verror(td, EIO, "write"); |
| 198 | |
| 199 | break; |
| 200 | } |
| 201 | } |
| 202 | |
| 203 | if (td->terminate) { |
| 204 | dprint(FD_FILE, "terminate unlink %s\n", f->file_name); |
| 205 | td_io_unlink_file(td, f); |
| 206 | } else if (td->o.create_fsync) { |
| 207 | if (fsync(f->fd) < 0) { |
| 208 | td_verror(td, errno, "fsync"); |
| 209 | goto err; |
| 210 | } |
| 211 | } |
| 212 | if (td->o.fill_device && !td_write(td)) { |
| 213 | fio_file_clear_size_known(f); |
| 214 | if (td_io_get_file_size(td, f)) |
| 215 | goto err; |
| 216 | if (f->io_size > f->real_file_size) |
| 217 | f->io_size = f->real_file_size; |
| 218 | } |
| 219 | |
| 220 | free(b); |
| 221 | done: |
| 222 | return 0; |
| 223 | err: |
| 224 | close(f->fd); |
| 225 | f->fd = -1; |
| 226 | if (b) |
| 227 | free(b); |
| 228 | return 1; |
| 229 | } |
| 230 | |
| 231 | static int pre_read_file(struct thread_data *td, struct fio_file *f) |
| 232 | { |
| 233 | int ret = 0, r, did_open = 0, old_runstate; |
| 234 | unsigned long long left; |
| 235 | unsigned int bs; |
| 236 | char *b; |
| 237 | |
| 238 | if (td_ioengine_flagged(td, FIO_PIPEIO) || |
| 239 | td_ioengine_flagged(td, FIO_NOIO)) |
| 240 | return 0; |
| 241 | |
| 242 | if (f->filetype == FIO_TYPE_CHAR) |
| 243 | return 0; |
| 244 | |
| 245 | if (!fio_file_open(f)) { |
| 246 | if (td->io_ops->open_file(td, f)) { |
| 247 | log_err("fio: cannot pre-read, failed to open file\n"); |
| 248 | return 1; |
| 249 | } |
| 250 | did_open = 1; |
| 251 | } |
| 252 | |
| 253 | old_runstate = td_bump_runstate(td, TD_PRE_READING); |
| 254 | |
| 255 | left = f->io_size; |
| 256 | bs = td->o.max_bs[DDIR_READ]; |
| 257 | if (bs > left) |
| 258 | bs = left; |
| 259 | |
| 260 | b = malloc(bs); |
| 261 | if (!b) { |
| 262 | td_verror(td, errno, "malloc"); |
| 263 | ret = 1; |
| 264 | goto error; |
| 265 | } |
| 266 | memset(b, 0, bs); |
| 267 | |
| 268 | if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) { |
| 269 | td_verror(td, errno, "lseek"); |
| 270 | log_err("fio: failed to lseek pre-read file\n"); |
| 271 | ret = 1; |
| 272 | goto error; |
| 273 | } |
| 274 | |
| 275 | while (left && !td->terminate) { |
| 276 | if (bs > left) |
| 277 | bs = left; |
| 278 | |
| 279 | r = read(f->fd, b, bs); |
| 280 | |
| 281 | if (r == (int) bs) { |
| 282 | left -= bs; |
| 283 | continue; |
| 284 | } else { |
| 285 | td_verror(td, EIO, "pre_read"); |
| 286 | break; |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | error: |
| 291 | td_restore_runstate(td, old_runstate); |
| 292 | |
| 293 | if (did_open) |
| 294 | td->io_ops->close_file(td, f); |
| 295 | |
| 296 | free(b); |
| 297 | return ret; |
| 298 | } |
| 299 | |
| 300 | unsigned long long get_rand_file_size(struct thread_data *td) |
| 301 | { |
| 302 | unsigned long long ret, sized; |
| 303 | uint64_t frand_max; |
| 304 | unsigned long r; |
| 305 | |
| 306 | frand_max = rand_max(&td->file_size_state); |
| 307 | r = __rand(&td->file_size_state); |
| 308 | sized = td->o.file_size_high - td->o.file_size_low; |
| 309 | ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0))); |
| 310 | ret += td->o.file_size_low; |
| 311 | ret -= (ret % td->o.rw_min_bs); |
| 312 | return ret; |
| 313 | } |
| 314 | |
| 315 | static int file_size(struct thread_data *td, struct fio_file *f) |
| 316 | { |
| 317 | struct stat st; |
| 318 | |
| 319 | if (stat(f->file_name, &st) == -1) { |
| 320 | td_verror(td, errno, "fstat"); |
| 321 | return 1; |
| 322 | } |
| 323 | |
| 324 | f->real_file_size = st.st_size; |
| 325 | return 0; |
| 326 | } |
| 327 | |
| 328 | static int bdev_size(struct thread_data *td, struct fio_file *f) |
| 329 | { |
| 330 | unsigned long long bytes = 0; |
| 331 | int r; |
| 332 | |
| 333 | if (td->io_ops->open_file(td, f)) { |
| 334 | log_err("fio: failed opening blockdev %s for size check\n", |
| 335 | f->file_name); |
| 336 | return 1; |
| 337 | } |
| 338 | |
| 339 | r = blockdev_size(f, &bytes); |
| 340 | if (r) { |
| 341 | td_verror(td, r, "blockdev_size"); |
| 342 | goto err; |
| 343 | } |
| 344 | |
| 345 | if (!bytes) { |
| 346 | log_err("%s: zero sized block device?\n", f->file_name); |
| 347 | goto err; |
| 348 | } |
| 349 | |
| 350 | f->real_file_size = bytes; |
| 351 | td->io_ops->close_file(td, f); |
| 352 | return 0; |
| 353 | err: |
| 354 | td->io_ops->close_file(td, f); |
| 355 | return 1; |
| 356 | } |
| 357 | |
| 358 | static int char_size(struct thread_data *td, struct fio_file *f) |
| 359 | { |
| 360 | #ifdef FIO_HAVE_CHARDEV_SIZE |
| 361 | unsigned long long bytes = 0; |
| 362 | int r; |
| 363 | |
| 364 | if (td->io_ops->open_file(td, f)) { |
| 365 | log_err("fio: failed opening chardev %s for size check\n", |
| 366 | f->file_name); |
| 367 | return 1; |
| 368 | } |
| 369 | |
| 370 | r = chardev_size(f, &bytes); |
| 371 | if (r) { |
| 372 | td_verror(td, r, "chardev_size"); |
| 373 | goto err; |
| 374 | } |
| 375 | |
| 376 | if (!bytes) { |
| 377 | log_err("%s: zero sized char device?\n", f->file_name); |
| 378 | goto err; |
| 379 | } |
| 380 | |
| 381 | f->real_file_size = bytes; |
| 382 | td->io_ops->close_file(td, f); |
| 383 | return 0; |
| 384 | err: |
| 385 | td->io_ops->close_file(td, f); |
| 386 | return 1; |
| 387 | #else |
| 388 | f->real_file_size = -1ULL; |
| 389 | return 0; |
| 390 | #endif |
| 391 | } |
| 392 | |
| 393 | static int get_file_size(struct thread_data *td, struct fio_file *f) |
| 394 | { |
| 395 | int ret = 0; |
| 396 | |
| 397 | if (fio_file_size_known(f)) |
| 398 | return 0; |
| 399 | |
| 400 | if (f->filetype == FIO_TYPE_FILE) |
| 401 | ret = file_size(td, f); |
| 402 | else if (f->filetype == FIO_TYPE_BLOCK) |
| 403 | ret = bdev_size(td, f); |
| 404 | else if (f->filetype == FIO_TYPE_CHAR) |
| 405 | ret = char_size(td, f); |
| 406 | else |
| 407 | f->real_file_size = -1ULL; |
| 408 | |
| 409 | /* |
| 410 | * Leave ->real_file_size with 0 since it could be expectation |
| 411 | * of initial setup for regular files. |
| 412 | */ |
| 413 | if (ret) |
| 414 | return ret; |
| 415 | |
| 416 | /* |
| 417 | * If ->real_file_size is -1, a conditional for the message |
| 418 | * "offset extends end" is always true, but it makes no sense, |
| 419 | * so just return the same value here. |
| 420 | */ |
| 421 | if (f->real_file_size == -1ULL) { |
| 422 | log_info("%s: failed to get file size of %s\n", td->o.name, |
| 423 | f->file_name); |
| 424 | return 1; |
| 425 | } |
| 426 | |
| 427 | if (td->o.start_offset && f->file_offset == 0) |
| 428 | dprint(FD_FILE, "offset of file %s not initialized yet\n", |
| 429 | f->file_name); |
| 430 | /* |
| 431 | * ->file_offset normally hasn't been initialized yet, so this |
| 432 | * is basically always false. |
| 433 | */ |
| 434 | if (f->file_offset > f->real_file_size) { |
| 435 | log_err("%s: offset extends end (%llu > %llu)\n", td->o.name, |
| 436 | (unsigned long long) f->file_offset, |
| 437 | (unsigned long long) f->real_file_size); |
| 438 | return 1; |
| 439 | } |
| 440 | |
| 441 | fio_file_set_size_known(f); |
| 442 | return 0; |
| 443 | } |
| 444 | |
| 445 | static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f, |
| 446 | unsigned long long off, |
| 447 | unsigned long long len) |
| 448 | { |
| 449 | int errval = 0, ret = 0; |
| 450 | |
| 451 | #ifdef CONFIG_ESX |
| 452 | return 0; |
| 453 | #endif |
| 454 | |
| 455 | if (len == -1ULL) |
| 456 | len = f->io_size; |
| 457 | if (off == -1ULL) |
| 458 | off = f->file_offset; |
| 459 | |
| 460 | if (len == -1ULL || off == -1ULL) |
| 461 | return 0; |
| 462 | |
| 463 | if (td->io_ops->invalidate) { |
| 464 | dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name, |
| 465 | f->file_name); |
| 466 | ret = td->io_ops->invalidate(td, f); |
| 467 | if (ret < 0) |
| 468 | errval = -ret; |
| 469 | } else if (f->filetype == FIO_TYPE_FILE) { |
| 470 | dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n", |
| 471 | f->file_name, off, len); |
| 472 | ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED); |
| 473 | if (ret) |
| 474 | errval = ret; |
| 475 | } else if (f->filetype == FIO_TYPE_BLOCK) { |
| 476 | int retry_count = 0; |
| 477 | |
| 478 | dprint(FD_IO, "drop page cache %s\n", f->file_name); |
| 479 | ret = blockdev_invalidate_cache(f); |
| 480 | while (ret < 0 && errno == EAGAIN && retry_count++ < 25) { |
| 481 | /* |
| 482 | * Linux multipath devices reject ioctl while |
| 483 | * the maps are being updated. That window can |
| 484 | * last tens of milliseconds; we'll try up to |
| 485 | * a quarter of a second. |
| 486 | */ |
| 487 | usleep(10000); |
| 488 | ret = blockdev_invalidate_cache(f); |
| 489 | } |
| 490 | if (ret < 0 && errno == EACCES && geteuid()) { |
| 491 | if (!root_warn) { |
| 492 | log_err("fio: only root may flush block " |
| 493 | "devices. Cache flush bypassed!\n"); |
| 494 | root_warn = 1; |
| 495 | } |
| 496 | ret = 0; |
| 497 | } |
| 498 | if (ret < 0) |
| 499 | errval = errno; |
| 500 | else if (ret) /* probably not supported */ |
| 501 | errval = ret; |
| 502 | } else if (f->filetype == FIO_TYPE_CHAR || |
| 503 | f->filetype == FIO_TYPE_PIPE) { |
| 504 | dprint(FD_IO, "invalidate not supported %s\n", f->file_name); |
| 505 | ret = 0; |
| 506 | } |
| 507 | |
| 508 | /* |
| 509 | * Cache flushing isn't a fatal condition, and we know it will |
| 510 | * happen on some platforms where we don't have the proper |
| 511 | * function to flush eg block device caches. So just warn and |
| 512 | * continue on our way. |
| 513 | */ |
| 514 | if (errval) |
| 515 | log_info("fio: cache invalidation of %s failed: %s\n", |
| 516 | f->file_name, strerror(errval)); |
| 517 | |
| 518 | return 0; |
| 519 | |
| 520 | } |
| 521 | |
| 522 | int file_invalidate_cache(struct thread_data *td, struct fio_file *f) |
| 523 | { |
| 524 | if (!fio_file_open(f)) |
| 525 | return 0; |
| 526 | |
| 527 | return __file_invalidate_cache(td, f, -1ULL, -1ULL); |
| 528 | } |
| 529 | |
| 530 | int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f) |
| 531 | { |
| 532 | int ret = 0; |
| 533 | |
| 534 | dprint(FD_FILE, "fd close %s\n", f->file_name); |
| 535 | |
| 536 | remove_file_hash(f); |
| 537 | |
| 538 | if (close(f->fd) < 0) |
| 539 | ret = errno; |
| 540 | |
| 541 | f->fd = -1; |
| 542 | |
| 543 | if (f->shadow_fd != -1) { |
| 544 | close(f->shadow_fd); |
| 545 | f->shadow_fd = -1; |
| 546 | } |
| 547 | |
| 548 | f->engine_pos = 0; |
| 549 | return ret; |
| 550 | } |
| 551 | |
| 552 | int file_lookup_open(struct fio_file *f, int flags) |
| 553 | { |
| 554 | struct fio_file *__f; |
| 555 | int from_hash; |
| 556 | |
| 557 | __f = lookup_file_hash(f->file_name); |
| 558 | if (__f) { |
| 559 | dprint(FD_FILE, "found file in hash %s\n", f->file_name); |
| 560 | f->lock = __f->lock; |
| 561 | from_hash = 1; |
| 562 | } else { |
| 563 | dprint(FD_FILE, "file not found in hash %s\n", f->file_name); |
| 564 | from_hash = 0; |
| 565 | } |
| 566 | |
| 567 | #ifdef WIN32 |
| 568 | flags |= _O_BINARY; |
| 569 | #endif |
| 570 | |
| 571 | f->fd = open(f->file_name, flags, 0600); |
| 572 | return from_hash; |
| 573 | } |
| 574 | |
| 575 | static int file_close_shadow_fds(struct thread_data *td) |
| 576 | { |
| 577 | struct fio_file *f; |
| 578 | int num_closed = 0; |
| 579 | unsigned int i; |
| 580 | |
| 581 | for_each_file(td, f, i) { |
| 582 | if (f->shadow_fd == -1) |
| 583 | continue; |
| 584 | |
| 585 | close(f->shadow_fd); |
| 586 | f->shadow_fd = -1; |
| 587 | num_closed++; |
| 588 | } |
| 589 | |
| 590 | return num_closed; |
| 591 | } |
| 592 | |
| 593 | int generic_open_file(struct thread_data *td, struct fio_file *f) |
| 594 | { |
| 595 | int is_std = 0; |
| 596 | int flags = 0; |
| 597 | int from_hash = 0; |
| 598 | |
| 599 | dprint(FD_FILE, "fd open %s\n", f->file_name); |
| 600 | |
| 601 | if (!strcmp(f->file_name, "-")) { |
| 602 | if (td_rw(td)) { |
| 603 | log_err("fio: can't read/write to stdin/out\n"); |
| 604 | return 1; |
| 605 | } |
| 606 | is_std = 1; |
| 607 | |
| 608 | /* |
| 609 | * move output logging to stderr, if we are writing to stdout |
| 610 | */ |
| 611 | if (td_write(td)) |
| 612 | f_out = stderr; |
| 613 | } |
| 614 | |
| 615 | if (td_trim(td)) |
| 616 | goto skip_flags; |
| 617 | if (td->o.odirect) |
| 618 | flags |= OS_O_DIRECT; |
| 619 | if (td->o.oatomic) { |
| 620 | if (!FIO_O_ATOMIC) { |
| 621 | td_verror(td, EINVAL, "OS does not support atomic IO"); |
| 622 | return 1; |
| 623 | } |
| 624 | flags |= OS_O_DIRECT | FIO_O_ATOMIC; |
| 625 | } |
| 626 | if (td->o.sync_io) |
| 627 | flags |= O_SYNC; |
| 628 | if (td->o.create_on_open && td->o.allow_create) |
| 629 | flags |= O_CREAT; |
| 630 | skip_flags: |
| 631 | if (f->filetype != FIO_TYPE_FILE) |
| 632 | flags |= FIO_O_NOATIME; |
| 633 | |
| 634 | open_again: |
| 635 | if (td_write(td)) { |
| 636 | if (!read_only) |
| 637 | flags |= O_RDWR; |
| 638 | |
| 639 | if (f->filetype == FIO_TYPE_FILE && td->o.allow_create) |
| 640 | flags |= O_CREAT; |
| 641 | |
| 642 | if (is_std) |
| 643 | f->fd = dup(STDOUT_FILENO); |
| 644 | else |
| 645 | from_hash = file_lookup_open(f, flags); |
| 646 | } else if (td_read(td)) { |
| 647 | if (f->filetype == FIO_TYPE_CHAR && !read_only) |
| 648 | flags |= O_RDWR; |
| 649 | else |
| 650 | flags |= O_RDONLY; |
| 651 | |
| 652 | if (is_std) |
| 653 | f->fd = dup(STDIN_FILENO); |
| 654 | else |
| 655 | from_hash = file_lookup_open(f, flags); |
| 656 | } else if (td_trim(td)) { |
| 657 | assert(!td_rw(td)); /* should have matched above */ |
| 658 | flags |= O_RDWR; |
| 659 | from_hash = file_lookup_open(f, flags); |
| 660 | } |
| 661 | |
| 662 | if (f->fd == -1) { |
| 663 | char buf[FIO_VERROR_SIZE]; |
| 664 | int __e = errno; |
| 665 | |
| 666 | if (__e == EPERM && (flags & FIO_O_NOATIME)) { |
| 667 | flags &= ~FIO_O_NOATIME; |
| 668 | goto open_again; |
| 669 | } |
| 670 | if (__e == EMFILE && file_close_shadow_fds(td)) |
| 671 | goto open_again; |
| 672 | |
| 673 | snprintf(buf, sizeof(buf), "open(%s)", f->file_name); |
| 674 | |
| 675 | if (__e == EINVAL && (flags & OS_O_DIRECT)) { |
| 676 | log_err("fio: looks like your file system does not " \ |
| 677 | "support direct=1/buffered=0\n"); |
| 678 | } |
| 679 | |
| 680 | td_verror(td, __e, buf); |
| 681 | return 1; |
| 682 | } |
| 683 | |
| 684 | if (!from_hash && f->fd != -1) { |
| 685 | if (add_file_hash(f)) { |
| 686 | int fio_unused ret; |
| 687 | |
| 688 | /* |
| 689 | * Stash away descriptor for later close. This is to |
| 690 | * work-around a "feature" on Linux, where a close of |
| 691 | * an fd that has been opened for write will trigger |
| 692 | * udev to call blkid to check partitions, fs id, etc. |
| 693 | * That pollutes the device cache, which can slow down |
| 694 | * unbuffered accesses. |
| 695 | */ |
| 696 | if (f->shadow_fd == -1) |
| 697 | f->shadow_fd = f->fd; |
| 698 | else { |
| 699 | /* |
| 700 | * OK to ignore, we haven't done anything |
| 701 | * with it |
| 702 | */ |
| 703 | ret = generic_close_file(td, f); |
| 704 | } |
| 705 | goto open_again; |
| 706 | } |
| 707 | } |
| 708 | |
| 709 | return 0; |
| 710 | } |
| 711 | |
| 712 | /* |
| 713 | * This function i.e. get_file_size() is the default .get_file_size |
| 714 | * implementation of majority of I/O engines. |
| 715 | */ |
| 716 | int generic_get_file_size(struct thread_data *td, struct fio_file *f) |
| 717 | { |
| 718 | return get_file_size(td, f); |
| 719 | } |
| 720 | |
| 721 | /* |
| 722 | * open/close all files, so that ->real_file_size gets set |
| 723 | */ |
| 724 | static int get_file_sizes(struct thread_data *td) |
| 725 | { |
| 726 | struct fio_file *f; |
| 727 | unsigned int i; |
| 728 | int err = 0; |
| 729 | |
| 730 | for_each_file(td, f, i) { |
| 731 | dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i, |
| 732 | f->file_name); |
| 733 | |
| 734 | if (td_io_get_file_size(td, f)) { |
| 735 | if (td->error != ENOENT) { |
| 736 | log_err("%s\n", td->verror); |
| 737 | err = 1; |
| 738 | break; |
| 739 | } |
| 740 | clear_error(td); |
| 741 | } |
| 742 | |
| 743 | /* |
| 744 | * There are corner cases where we end up with -1 for |
| 745 | * ->real_file_size due to unsupported file type, etc. |
| 746 | * We then just set to size option value divided by number |
| 747 | * of files, similar to the way file ->io_size is set. |
| 748 | * stat(2) failure doesn't set ->real_file_size to -1. |
| 749 | */ |
| 750 | if (f->real_file_size == -1ULL && td->o.size) |
| 751 | f->real_file_size = td->o.size / td->o.nr_files; |
| 752 | } |
| 753 | |
| 754 | return err; |
| 755 | } |
| 756 | |
| 757 | struct fio_mount { |
| 758 | struct flist_head list; |
| 759 | const char *base; |
| 760 | char __base[256]; |
| 761 | unsigned int key; |
| 762 | }; |
| 763 | |
| 764 | /* |
| 765 | * Get free number of bytes for each file on each unique mount. |
| 766 | */ |
| 767 | static unsigned long long get_fs_free_counts(struct thread_data *td) |
| 768 | { |
| 769 | struct flist_head *n, *tmp; |
| 770 | unsigned long long ret = 0; |
| 771 | struct fio_mount *fm; |
| 772 | FLIST_HEAD(list); |
| 773 | struct fio_file *f; |
| 774 | unsigned int i; |
| 775 | |
| 776 | for_each_file(td, f, i) { |
| 777 | struct stat sb; |
| 778 | char buf[256]; |
| 779 | |
| 780 | if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) { |
| 781 | if (f->real_file_size != -1ULL) |
| 782 | ret += f->real_file_size; |
| 783 | continue; |
| 784 | } else if (f->filetype != FIO_TYPE_FILE) |
| 785 | continue; |
| 786 | |
| 787 | buf[255] = '\0'; |
| 788 | strncpy(buf, f->file_name, 255); |
| 789 | |
| 790 | if (stat(buf, &sb) < 0) { |
| 791 | if (errno != ENOENT) |
| 792 | break; |
| 793 | strcpy(buf, "."); |
| 794 | if (stat(buf, &sb) < 0) |
| 795 | break; |
| 796 | } |
| 797 | |
| 798 | fm = NULL; |
| 799 | flist_for_each(n, &list) { |
| 800 | fm = flist_entry(n, struct fio_mount, list); |
| 801 | if (fm->key == sb.st_dev) |
| 802 | break; |
| 803 | |
| 804 | fm = NULL; |
| 805 | } |
| 806 | |
| 807 | if (fm) |
| 808 | continue; |
| 809 | |
| 810 | fm = calloc(1, sizeof(*fm)); |
| 811 | strncpy(fm->__base, buf, sizeof(fm->__base) - 1); |
| 812 | fm->base = basename(fm->__base); |
| 813 | fm->key = sb.st_dev; |
| 814 | flist_add(&fm->list, &list); |
| 815 | } |
| 816 | |
| 817 | flist_for_each_safe(n, tmp, &list) { |
| 818 | unsigned long long sz; |
| 819 | |
| 820 | fm = flist_entry(n, struct fio_mount, list); |
| 821 | flist_del(&fm->list); |
| 822 | |
| 823 | sz = get_fs_free_size(fm->base); |
| 824 | if (sz && sz != -1ULL) |
| 825 | ret += sz; |
| 826 | |
| 827 | free(fm); |
| 828 | } |
| 829 | |
| 830 | return ret; |
| 831 | } |
| 832 | |
| 833 | uint64_t get_start_offset(struct thread_data *td, struct fio_file *f) |
| 834 | { |
| 835 | struct thread_options *o = &td->o; |
| 836 | unsigned long long align_bs; |
| 837 | unsigned long long offset; |
| 838 | |
| 839 | if (o->file_append && f->filetype == FIO_TYPE_FILE) |
| 840 | return f->real_file_size; |
| 841 | |
| 842 | if (o->start_offset_percent > 0) { |
| 843 | /* |
| 844 | * if blockalign is provided, find the min across read, write, |
| 845 | * and trim |
| 846 | */ |
| 847 | if (fio_option_is_set(o, ba)) { |
| 848 | align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]); |
| 849 | align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs); |
| 850 | } else { |
| 851 | /* else take the minimum block size */ |
| 852 | align_bs = td_min_bs(td); |
| 853 | } |
| 854 | |
| 855 | /* calculate the raw offset */ |
| 856 | offset = (f->real_file_size * o->start_offset_percent / 100) + |
| 857 | (td->subjob_number * o->offset_increment); |
| 858 | |
| 859 | /* |
| 860 | * block align the offset at the next available boundary at |
| 861 | * ceiling(offset / align_bs) * align_bs |
| 862 | */ |
| 863 | offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs; |
| 864 | |
| 865 | } else { |
| 866 | /* start_offset_percent not set */ |
| 867 | offset = o->start_offset + |
| 868 | td->subjob_number * o->offset_increment; |
| 869 | } |
| 870 | |
| 871 | return offset; |
| 872 | } |
| 873 | |
| 874 | /* |
| 875 | * Open the files and setup files sizes, creating files if necessary. |
| 876 | */ |
| 877 | int setup_files(struct thread_data *td) |
| 878 | { |
| 879 | unsigned long long total_size, extend_size; |
| 880 | struct thread_options *o = &td->o; |
| 881 | struct fio_file *f; |
| 882 | unsigned int i, nr_fs_extra = 0; |
| 883 | int err = 0, need_extend; |
| 884 | int old_state; |
| 885 | const unsigned int bs = td_min_bs(td); |
| 886 | uint64_t fs = 0; |
| 887 | |
| 888 | dprint(FD_FILE, "setup files\n"); |
| 889 | |
| 890 | old_state = td_bump_runstate(td, TD_SETTING_UP); |
| 891 | |
| 892 | if (o->read_iolog_file) |
| 893 | goto done; |
| 894 | |
| 895 | /* |
| 896 | * Find out physical size of files or devices for this thread, |
| 897 | * before we determine I/O size and range of our targets. |
| 898 | * If ioengine defines a setup() method, it's responsible for |
| 899 | * opening the files and setting f->real_file_size to indicate |
| 900 | * the valid range for that file. |
| 901 | */ |
| 902 | if (td->io_ops->setup) |
| 903 | err = td->io_ops->setup(td); |
| 904 | else |
| 905 | err = get_file_sizes(td); |
| 906 | |
| 907 | if (err) |
| 908 | goto err_out; |
| 909 | |
| 910 | /* |
| 911 | * check sizes. if the files/devices do not exist and the size |
| 912 | * isn't passed to fio, abort. |
| 913 | */ |
| 914 | total_size = 0; |
| 915 | for_each_file(td, f, i) { |
| 916 | f->fileno = i; |
| 917 | if (f->real_file_size == -1ULL) |
| 918 | total_size = -1ULL; |
| 919 | else |
| 920 | total_size += f->real_file_size; |
| 921 | } |
| 922 | |
| 923 | if (o->fill_device) |
| 924 | td->fill_device_size = get_fs_free_counts(td); |
| 925 | |
| 926 | /* |
| 927 | * device/file sizes are zero and no size given, punt |
| 928 | */ |
| 929 | if ((!total_size || total_size == -1ULL) && !o->size && |
| 930 | !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device && |
| 931 | !(o->nr_files && (o->file_size_low || o->file_size_high))) { |
| 932 | log_err("%s: you need to specify size=\n", o->name); |
| 933 | td_verror(td, EINVAL, "total_file_size"); |
| 934 | goto err_out; |
| 935 | } |
| 936 | |
| 937 | /* |
| 938 | * Calculate per-file size and potential extra size for the |
| 939 | * first files, if needed (i.e. if we don't have a fixed size). |
| 940 | */ |
| 941 | if (!o->file_size_low && o->nr_files) { |
| 942 | uint64_t all_fs; |
| 943 | |
| 944 | fs = o->size / o->nr_files; |
| 945 | all_fs = fs * o->nr_files; |
| 946 | |
| 947 | if (all_fs < o->size) |
| 948 | nr_fs_extra = (o->size - all_fs) / bs; |
| 949 | } |
| 950 | |
| 951 | /* |
| 952 | * now file sizes are known, so we can set ->io_size. if size= is |
| 953 | * not given, ->io_size is just equal to ->real_file_size. if size |
| 954 | * is given, ->io_size is size / nr_files. |
| 955 | */ |
| 956 | extend_size = total_size = 0; |
| 957 | need_extend = 0; |
| 958 | for_each_file(td, f, i) { |
| 959 | f->file_offset = get_start_offset(td, f); |
| 960 | |
| 961 | /* |
| 962 | * Update ->io_size depending on options specified. |
| 963 | * ->file_size_low being 0 means filesize option isn't set. |
| 964 | * Non zero ->file_size_low equals ->file_size_high means |
| 965 | * filesize option is set in a fixed size format. |
| 966 | * Non zero ->file_size_low not equals ->file_size_high means |
| 967 | * filesize option is set in a range format. |
| 968 | */ |
| 969 | if (!o->file_size_low) { |
| 970 | /* |
| 971 | * no file size or range given, file size is equal to |
| 972 | * total size divided by number of files. If the size |
| 973 | * doesn't divide nicely with the min blocksize, |
| 974 | * make the first files bigger. |
| 975 | */ |
| 976 | f->io_size = fs; |
| 977 | if (nr_fs_extra) { |
| 978 | nr_fs_extra--; |
| 979 | f->io_size += bs; |
| 980 | } |
| 981 | |
| 982 | /* |
| 983 | * We normally don't come here for regular files, but |
| 984 | * if the result is 0 for a regular file, set it to the |
| 985 | * real file size. This could be size of the existing |
| 986 | * one if it already exists, but otherwise will be set |
| 987 | * to 0. A new file won't be created because |
| 988 | * ->io_size + ->file_offset equals ->real_file_size. |
| 989 | */ |
| 990 | if (!f->io_size) { |
| 991 | if (f->file_offset > f->real_file_size) |
| 992 | goto err_offset; |
| 993 | f->io_size = f->real_file_size - f->file_offset; |
| 994 | if (!f->io_size) |
| 995 | log_info("fio: file %s may be ignored\n", |
| 996 | f->file_name); |
| 997 | } |
| 998 | } else if (f->real_file_size < o->file_size_low || |
| 999 | f->real_file_size > o->file_size_high) { |
| 1000 | if (f->file_offset > o->file_size_low) |
| 1001 | goto err_offset; |
| 1002 | /* |
| 1003 | * file size given. if it's fixed, use that. if it's a |
| 1004 | * range, generate a random size in-between. |
| 1005 | */ |
| 1006 | if (o->file_size_low == o->file_size_high) |
| 1007 | f->io_size = o->file_size_low - f->file_offset; |
| 1008 | else { |
| 1009 | f->io_size = get_rand_file_size(td) |
| 1010 | - f->file_offset; |
| 1011 | } |
| 1012 | } else |
| 1013 | f->io_size = f->real_file_size - f->file_offset; |
| 1014 | |
| 1015 | if (f->io_size == -1ULL) |
| 1016 | total_size = -1ULL; |
| 1017 | else { |
| 1018 | if (o->size_percent) { |
| 1019 | uint64_t file_size; |
| 1020 | |
| 1021 | file_size = f->io_size + f->file_offset; |
| 1022 | f->io_size = (file_size * |
| 1023 | o->size_percent) / 100; |
| 1024 | if (f->io_size > (file_size - f->file_offset)) |
| 1025 | f->io_size = file_size - f->file_offset; |
| 1026 | |
| 1027 | f->io_size -= (f->io_size % td_min_bs(td)); |
| 1028 | } |
| 1029 | total_size += f->io_size; |
| 1030 | } |
| 1031 | |
| 1032 | if (f->filetype == FIO_TYPE_FILE && |
| 1033 | (f->io_size + f->file_offset) > f->real_file_size && |
| 1034 | !td_ioengine_flagged(td, FIO_DISKLESSIO)) { |
| 1035 | if (!o->create_on_open) { |
| 1036 | need_extend++; |
| 1037 | extend_size += (f->io_size + f->file_offset); |
| 1038 | fio_file_set_extend(f); |
| 1039 | } else |
| 1040 | f->real_file_size = f->io_size + f->file_offset; |
| 1041 | } |
| 1042 | } |
| 1043 | |
| 1044 | if (td->o.block_error_hist) { |
| 1045 | int len; |
| 1046 | |
| 1047 | assert(td->o.nr_files == 1); /* checked in fixup_options */ |
| 1048 | f = td->files[0]; |
| 1049 | len = f->io_size / td->o.bs[DDIR_TRIM]; |
| 1050 | if (len > MAX_NR_BLOCK_INFOS || len <= 0) { |
| 1051 | log_err("fio: cannot calculate block histogram with " |
| 1052 | "%d trim blocks, maximum %d\n", |
| 1053 | len, MAX_NR_BLOCK_INFOS); |
| 1054 | td_verror(td, EINVAL, "block_error_hist"); |
| 1055 | goto err_out; |
| 1056 | } |
| 1057 | |
| 1058 | td->ts.nr_block_infos = len; |
| 1059 | for (i = 0; i < len; i++) |
| 1060 | td->ts.block_infos[i] = |
| 1061 | BLOCK_INFO(0, BLOCK_STATE_UNINIT); |
| 1062 | } else |
| 1063 | td->ts.nr_block_infos = 0; |
| 1064 | |
| 1065 | if (!o->size || (total_size && o->size > total_size)) |
| 1066 | o->size = total_size; |
| 1067 | |
| 1068 | if (o->size < td_min_bs(td)) { |
| 1069 | log_err("fio: blocksize too large for data set\n"); |
| 1070 | goto err_out; |
| 1071 | } |
| 1072 | |
| 1073 | /* |
| 1074 | * See if we need to extend some files, typically needed when our |
| 1075 | * target regular files don't exist yet, but our jobs require them |
| 1076 | * initially due to read I/Os. |
| 1077 | */ |
| 1078 | if (need_extend) { |
| 1079 | temp_stall_ts = 1; |
| 1080 | if (output_format & FIO_OUTPUT_NORMAL) { |
| 1081 | log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n", |
| 1082 | o->name, |
| 1083 | need_extend > 1 ? "s" : "", |
| 1084 | need_extend, |
| 1085 | need_extend > 1 ? "s" : "", |
| 1086 | need_extend > 1 ? "total " : "", |
| 1087 | extend_size >> 20); |
| 1088 | } |
| 1089 | |
| 1090 | for_each_file(td, f, i) { |
| 1091 | unsigned long long old_len = -1ULL, extend_len = -1ULL; |
| 1092 | |
| 1093 | if (!fio_file_extend(f)) |
| 1094 | continue; |
| 1095 | |
| 1096 | assert(f->filetype == FIO_TYPE_FILE); |
| 1097 | fio_file_clear_extend(f); |
| 1098 | if (!o->fill_device) { |
| 1099 | old_len = f->real_file_size; |
| 1100 | extend_len = f->io_size + f->file_offset - |
| 1101 | old_len; |
| 1102 | } |
| 1103 | f->real_file_size = (f->io_size + f->file_offset); |
| 1104 | err = extend_file(td, f); |
| 1105 | if (err) |
| 1106 | break; |
| 1107 | |
| 1108 | err = __file_invalidate_cache(td, f, old_len, |
| 1109 | extend_len); |
| 1110 | |
| 1111 | /* |
| 1112 | * Shut up static checker |
| 1113 | */ |
| 1114 | if (f->fd != -1) |
| 1115 | close(f->fd); |
| 1116 | |
| 1117 | f->fd = -1; |
| 1118 | if (err) |
| 1119 | break; |
| 1120 | } |
| 1121 | temp_stall_ts = 0; |
| 1122 | } |
| 1123 | |
| 1124 | if (err) |
| 1125 | goto err_out; |
| 1126 | |
| 1127 | if (!o->zone_size) |
| 1128 | o->zone_size = o->size; |
| 1129 | |
| 1130 | /* |
| 1131 | * iolog already set the total io size, if we read back |
| 1132 | * stored entries. |
| 1133 | */ |
| 1134 | if (!o->read_iolog_file) { |
| 1135 | if (o->io_size) |
| 1136 | td->total_io_size = o->io_size * o->loops; |
| 1137 | else |
| 1138 | td->total_io_size = o->size * o->loops; |
| 1139 | } |
| 1140 | |
| 1141 | done: |
| 1142 | if (o->create_only) |
| 1143 | td->done = 1; |
| 1144 | |
| 1145 | td_restore_runstate(td, old_state); |
| 1146 | return 0; |
| 1147 | err_offset: |
| 1148 | log_err("%s: you need to specify valid offset=\n", o->name); |
| 1149 | err_out: |
| 1150 | td_restore_runstate(td, old_state); |
| 1151 | return 1; |
| 1152 | } |
| 1153 | |
| 1154 | int pre_read_files(struct thread_data *td) |
| 1155 | { |
| 1156 | struct fio_file *f; |
| 1157 | unsigned int i; |
| 1158 | |
| 1159 | dprint(FD_FILE, "pre_read files\n"); |
| 1160 | |
| 1161 | for_each_file(td, f, i) { |
| 1162 | if (pre_read_file(td, f)) |
| 1163 | return -1; |
| 1164 | } |
| 1165 | |
| 1166 | return 0; |
| 1167 | } |
| 1168 | |
| 1169 | static int __init_rand_distribution(struct thread_data *td, struct fio_file *f) |
| 1170 | { |
| 1171 | unsigned int range_size, seed; |
| 1172 | unsigned long nranges; |
| 1173 | uint64_t fsize; |
| 1174 | |
| 1175 | range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]); |
| 1176 | fsize = min(f->real_file_size, f->io_size); |
| 1177 | |
| 1178 | nranges = (fsize + range_size - 1) / range_size; |
| 1179 | |
| 1180 | seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number; |
| 1181 | if (!td->o.rand_repeatable) |
| 1182 | seed = td->rand_seeds[4]; |
| 1183 | |
| 1184 | if (td->o.random_distribution == FIO_RAND_DIST_ZIPF) |
| 1185 | zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed); |
| 1186 | else if (td->o.random_distribution == FIO_RAND_DIST_PARETO) |
| 1187 | pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed); |
| 1188 | else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS) |
| 1189 | gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed); |
| 1190 | |
| 1191 | return 1; |
| 1192 | } |
| 1193 | |
| 1194 | static int init_rand_distribution(struct thread_data *td) |
| 1195 | { |
| 1196 | struct fio_file *f; |
| 1197 | unsigned int i; |
| 1198 | int state; |
| 1199 | |
| 1200 | if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) |
| 1201 | return 0; |
| 1202 | |
| 1203 | state = td_bump_runstate(td, TD_SETTING_UP); |
| 1204 | |
| 1205 | for_each_file(td, f, i) |
| 1206 | __init_rand_distribution(td, f); |
| 1207 | |
| 1208 | td_restore_runstate(td, state); |
| 1209 | |
| 1210 | return 1; |
| 1211 | } |
| 1212 | |
| 1213 | /* |
| 1214 | * Check if the number of blocks exceeds the randomness capability of |
| 1215 | * the selected generator. Tausworthe is 32-bit, the others are fullly |
| 1216 | * 64-bit capable. |
| 1217 | */ |
| 1218 | static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f, |
| 1219 | uint64_t blocks) |
| 1220 | { |
| 1221 | if (blocks <= FRAND32_MAX) |
| 1222 | return 0; |
| 1223 | if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE) |
| 1224 | return 0; |
| 1225 | |
| 1226 | /* |
| 1227 | * If the user hasn't specified a random generator, switch |
| 1228 | * to tausworthe64 with informational warning. If the user did |
| 1229 | * specify one, just warn. |
| 1230 | */ |
| 1231 | log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n", |
| 1232 | f->file_name); |
| 1233 | |
| 1234 | if (!fio_option_is_set(&td->o, random_generator)) { |
| 1235 | log_info("fio: Switching to tausworthe64. Use the " |
| 1236 | "random_generator= option to get rid of this " |
| 1237 | "warning.\n"); |
| 1238 | td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64; |
| 1239 | return 0; |
| 1240 | } |
| 1241 | |
| 1242 | /* |
| 1243 | * Just make this information to avoid breaking scripts. |
| 1244 | */ |
| 1245 | log_info("fio: Use the random_generator= option to switch to lfsr or " |
| 1246 | "tausworthe64.\n"); |
| 1247 | return 0; |
| 1248 | } |
| 1249 | |
| 1250 | int init_random_map(struct thread_data *td) |
| 1251 | { |
| 1252 | unsigned long long blocks; |
| 1253 | struct fio_file *f; |
| 1254 | unsigned int i; |
| 1255 | |
| 1256 | if (init_rand_distribution(td)) |
| 1257 | return 0; |
| 1258 | if (!td_random(td)) |
| 1259 | return 0; |
| 1260 | |
| 1261 | for_each_file(td, f, i) { |
| 1262 | uint64_t fsize = min(f->real_file_size, f->io_size); |
| 1263 | |
| 1264 | blocks = fsize / (unsigned long long) td->o.rw_min_bs; |
| 1265 | |
| 1266 | if (check_rand_gen_limits(td, f, blocks)) |
| 1267 | return 1; |
| 1268 | |
| 1269 | if (td->o.random_generator == FIO_RAND_GEN_LFSR) { |
| 1270 | unsigned long seed; |
| 1271 | |
| 1272 | seed = td->rand_seeds[FIO_RAND_BLOCK_OFF]; |
| 1273 | |
| 1274 | if (!lfsr_init(&f->lfsr, blocks, seed, 0)) { |
| 1275 | fio_file_set_lfsr(f); |
| 1276 | continue; |
| 1277 | } |
| 1278 | } else if (!td->o.norandommap) { |
| 1279 | f->io_axmap = axmap_new(blocks); |
| 1280 | if (f->io_axmap) { |
| 1281 | fio_file_set_axmap(f); |
| 1282 | continue; |
| 1283 | } |
| 1284 | } else if (td->o.norandommap) |
| 1285 | continue; |
| 1286 | |
| 1287 | if (!td->o.softrandommap) { |
| 1288 | log_err("fio: failed allocating random map. If running" |
| 1289 | " a large number of jobs, try the 'norandommap'" |
| 1290 | " option or set 'softrandommap'. Or give" |
| 1291 | " a larger --alloc-size to fio.\n"); |
| 1292 | return 1; |
| 1293 | } |
| 1294 | |
| 1295 | log_info("fio: file %s failed allocating random map. Running " |
| 1296 | "job without.\n", f->file_name); |
| 1297 | } |
| 1298 | |
| 1299 | return 0; |
| 1300 | } |
| 1301 | |
| 1302 | void close_files(struct thread_data *td) |
| 1303 | { |
| 1304 | struct fio_file *f; |
| 1305 | unsigned int i; |
| 1306 | |
| 1307 | for_each_file(td, f, i) { |
| 1308 | if (fio_file_open(f)) |
| 1309 | td_io_close_file(td, f); |
| 1310 | } |
| 1311 | } |
| 1312 | |
| 1313 | void close_and_free_files(struct thread_data *td) |
| 1314 | { |
| 1315 | struct fio_file *f; |
| 1316 | unsigned int i; |
| 1317 | |
| 1318 | dprint(FD_FILE, "close files\n"); |
| 1319 | |
| 1320 | for_each_file(td, f, i) { |
| 1321 | if (td->o.unlink && f->filetype == FIO_TYPE_FILE) { |
| 1322 | dprint(FD_FILE, "free unlink %s\n", f->file_name); |
| 1323 | td_io_unlink_file(td, f); |
| 1324 | } |
| 1325 | |
| 1326 | if (fio_file_open(f)) |
| 1327 | td_io_close_file(td, f); |
| 1328 | |
| 1329 | remove_file_hash(f); |
| 1330 | |
| 1331 | if (td->o.unlink && f->filetype == FIO_TYPE_FILE) { |
| 1332 | dprint(FD_FILE, "free unlink %s\n", f->file_name); |
| 1333 | td_io_unlink_file(td, f); |
| 1334 | } |
| 1335 | |
| 1336 | sfree(f->file_name); |
| 1337 | f->file_name = NULL; |
| 1338 | if (fio_file_axmap(f)) { |
| 1339 | axmap_free(f->io_axmap); |
| 1340 | f->io_axmap = NULL; |
| 1341 | } |
| 1342 | sfree(f); |
| 1343 | } |
| 1344 | |
| 1345 | td->o.filename = NULL; |
| 1346 | free(td->files); |
| 1347 | free(td->file_locks); |
| 1348 | td->files_index = 0; |
| 1349 | td->files = NULL; |
| 1350 | td->file_locks = NULL; |
| 1351 | td->o.file_lock_mode = FILE_LOCK_NONE; |
| 1352 | td->o.nr_files = 0; |
| 1353 | } |
| 1354 | |
| 1355 | static void get_file_type(struct fio_file *f) |
| 1356 | { |
| 1357 | struct stat sb; |
| 1358 | |
| 1359 | if (!strcmp(f->file_name, "-")) |
| 1360 | f->filetype = FIO_TYPE_PIPE; |
| 1361 | else |
| 1362 | f->filetype = FIO_TYPE_FILE; |
| 1363 | |
| 1364 | #ifdef WIN32 |
| 1365 | /* \\.\ is the device namespace in Windows, where every file is |
| 1366 | * a block device */ |
| 1367 | if (strncmp(f->file_name, "\\\\.\\", 4) == 0) |
| 1368 | f->filetype = FIO_TYPE_BLOCK; |
| 1369 | #endif |
| 1370 | |
| 1371 | if (!stat(f->file_name, &sb)) { |
| 1372 | if (S_ISBLK(sb.st_mode)) |
| 1373 | f->filetype = FIO_TYPE_BLOCK; |
| 1374 | else if (S_ISCHR(sb.st_mode)) |
| 1375 | f->filetype = FIO_TYPE_CHAR; |
| 1376 | else if (S_ISFIFO(sb.st_mode)) |
| 1377 | f->filetype = FIO_TYPE_PIPE; |
| 1378 | } |
| 1379 | } |
| 1380 | |
| 1381 | static bool __is_already_allocated(const char *fname, bool set) |
| 1382 | { |
| 1383 | struct flist_head *entry; |
| 1384 | bool ret; |
| 1385 | |
| 1386 | ret = file_bloom_exists(fname, set); |
| 1387 | if (!ret) |
| 1388 | return ret; |
| 1389 | |
| 1390 | flist_for_each(entry, &filename_list) { |
| 1391 | struct file_name *fn; |
| 1392 | |
| 1393 | fn = flist_entry(entry, struct file_name, list); |
| 1394 | |
| 1395 | if (!strcmp(fn->filename, fname)) |
| 1396 | return true; |
| 1397 | } |
| 1398 | |
| 1399 | return false; |
| 1400 | } |
| 1401 | |
| 1402 | static bool is_already_allocated(const char *fname) |
| 1403 | { |
| 1404 | bool ret; |
| 1405 | |
| 1406 | fio_file_hash_lock(); |
| 1407 | ret = __is_already_allocated(fname, false); |
| 1408 | fio_file_hash_unlock(); |
| 1409 | |
| 1410 | return ret; |
| 1411 | } |
| 1412 | |
| 1413 | static void set_already_allocated(const char *fname) |
| 1414 | { |
| 1415 | struct file_name *fn; |
| 1416 | |
| 1417 | fn = malloc(sizeof(struct file_name)); |
| 1418 | fn->filename = strdup(fname); |
| 1419 | |
| 1420 | fio_file_hash_lock(); |
| 1421 | if (!__is_already_allocated(fname, true)) { |
| 1422 | flist_add_tail(&fn->list, &filename_list); |
| 1423 | fn = NULL; |
| 1424 | } |
| 1425 | fio_file_hash_unlock(); |
| 1426 | |
| 1427 | if (fn) { |
| 1428 | free(fn->filename); |
| 1429 | free(fn); |
| 1430 | } |
| 1431 | } |
| 1432 | |
| 1433 | static void free_already_allocated(void) |
| 1434 | { |
| 1435 | struct flist_head *entry, *tmp; |
| 1436 | struct file_name *fn; |
| 1437 | |
| 1438 | if (flist_empty(&filename_list)) |
| 1439 | return; |
| 1440 | |
| 1441 | fio_file_hash_lock(); |
| 1442 | flist_for_each_safe(entry, tmp, &filename_list) { |
| 1443 | fn = flist_entry(entry, struct file_name, list); |
| 1444 | free(fn->filename); |
| 1445 | flist_del(&fn->list); |
| 1446 | free(fn); |
| 1447 | } |
| 1448 | |
| 1449 | fio_file_hash_unlock(); |
| 1450 | } |
| 1451 | |
| 1452 | static struct fio_file *alloc_new_file(struct thread_data *td) |
| 1453 | { |
| 1454 | struct fio_file *f; |
| 1455 | |
| 1456 | f = smalloc(sizeof(*f)); |
| 1457 | if (!f) { |
| 1458 | assert(0); |
| 1459 | return NULL; |
| 1460 | } |
| 1461 | |
| 1462 | f->fd = -1; |
| 1463 | f->shadow_fd = -1; |
| 1464 | fio_file_reset(td, f); |
| 1465 | return f; |
| 1466 | } |
| 1467 | |
| 1468 | bool exists_and_not_regfile(const char *filename) |
| 1469 | { |
| 1470 | struct stat sb; |
| 1471 | |
| 1472 | if (lstat(filename, &sb) == -1) |
| 1473 | return false; |
| 1474 | |
| 1475 | #ifndef WIN32 /* NOT Windows */ |
| 1476 | if (S_ISREG(sb.st_mode)) |
| 1477 | return false; |
| 1478 | #else |
| 1479 | /* \\.\ is the device namespace in Windows, where every file |
| 1480 | * is a device node */ |
| 1481 | if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0) |
| 1482 | return false; |
| 1483 | #endif |
| 1484 | |
| 1485 | return true; |
| 1486 | } |
| 1487 | |
| 1488 | int add_file(struct thread_data *td, const char *fname, int numjob, int inc) |
| 1489 | { |
| 1490 | int cur_files = td->files_index; |
| 1491 | char file_name[PATH_MAX]; |
| 1492 | struct fio_file *f; |
| 1493 | int len = 0; |
| 1494 | |
| 1495 | dprint(FD_FILE, "add file %s\n", fname); |
| 1496 | |
| 1497 | if (td->o.directory) |
| 1498 | len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob, |
| 1499 | td->o.unique_filename); |
| 1500 | |
| 1501 | sprintf(file_name + len, "%s", fname); |
| 1502 | |
| 1503 | /* clean cloned siblings using existing files */ |
| 1504 | if (numjob && is_already_allocated(file_name) && |
| 1505 | !exists_and_not_regfile(fname)) |
| 1506 | return 0; |
| 1507 | |
| 1508 | f = alloc_new_file(td); |
| 1509 | |
| 1510 | if (td->files_size <= td->files_index) { |
| 1511 | unsigned int new_size = td->o.nr_files + 1; |
| 1512 | |
| 1513 | dprint(FD_FILE, "resize file array to %d files\n", new_size); |
| 1514 | |
| 1515 | td->files = realloc(td->files, new_size * sizeof(f)); |
| 1516 | if (td->files == NULL) { |
| 1517 | log_err("fio: realloc OOM\n"); |
| 1518 | assert(0); |
| 1519 | } |
| 1520 | if (td->o.file_lock_mode != FILE_LOCK_NONE) { |
| 1521 | td->file_locks = realloc(td->file_locks, new_size); |
| 1522 | if (!td->file_locks) { |
| 1523 | log_err("fio: realloc OOM\n"); |
| 1524 | assert(0); |
| 1525 | } |
| 1526 | td->file_locks[cur_files] = FILE_LOCK_NONE; |
| 1527 | } |
| 1528 | td->files_size = new_size; |
| 1529 | } |
| 1530 | td->files[cur_files] = f; |
| 1531 | f->fileno = cur_files; |
| 1532 | |
| 1533 | /* |
| 1534 | * init function, io engine may not be loaded yet |
| 1535 | */ |
| 1536 | if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO)) |
| 1537 | f->real_file_size = -1ULL; |
| 1538 | |
| 1539 | f->file_name = smalloc_strdup(file_name); |
| 1540 | if (!f->file_name) |
| 1541 | assert(0); |
| 1542 | |
| 1543 | get_file_type(f); |
| 1544 | |
| 1545 | switch (td->o.file_lock_mode) { |
| 1546 | case FILE_LOCK_NONE: |
| 1547 | break; |
| 1548 | case FILE_LOCK_READWRITE: |
| 1549 | f->rwlock = fio_rwlock_init(); |
| 1550 | break; |
| 1551 | case FILE_LOCK_EXCLUSIVE: |
| 1552 | f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED); |
| 1553 | break; |
| 1554 | default: |
| 1555 | log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode); |
| 1556 | assert(0); |
| 1557 | } |
| 1558 | |
| 1559 | td->files_index++; |
| 1560 | if (f->filetype == FIO_TYPE_FILE) |
| 1561 | td->nr_normal_files++; |
| 1562 | |
| 1563 | set_already_allocated(file_name); |
| 1564 | |
| 1565 | if (inc) |
| 1566 | td->o.nr_files++; |
| 1567 | |
| 1568 | dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name, |
| 1569 | cur_files); |
| 1570 | |
| 1571 | return cur_files; |
| 1572 | } |
| 1573 | |
| 1574 | int add_file_exclusive(struct thread_data *td, const char *fname) |
| 1575 | { |
| 1576 | struct fio_file *f; |
| 1577 | unsigned int i; |
| 1578 | |
| 1579 | for_each_file(td, f, i) { |
| 1580 | if (!strcmp(f->file_name, fname)) |
| 1581 | return i; |
| 1582 | } |
| 1583 | |
| 1584 | return add_file(td, fname, 0, 1); |
| 1585 | } |
| 1586 | |
| 1587 | void get_file(struct fio_file *f) |
| 1588 | { |
| 1589 | dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references); |
| 1590 | assert(fio_file_open(f)); |
| 1591 | f->references++; |
| 1592 | } |
| 1593 | |
| 1594 | int put_file(struct thread_data *td, struct fio_file *f) |
| 1595 | { |
| 1596 | int f_ret = 0, ret = 0; |
| 1597 | |
| 1598 | dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references); |
| 1599 | |
| 1600 | if (!fio_file_open(f)) { |
| 1601 | assert(f->fd == -1); |
| 1602 | return 0; |
| 1603 | } |
| 1604 | |
| 1605 | assert(f->references); |
| 1606 | if (--f->references) |
| 1607 | return 0; |
| 1608 | |
| 1609 | if (should_fsync(td) && td->o.fsync_on_close) { |
| 1610 | f_ret = fsync(f->fd); |
| 1611 | if (f_ret < 0) |
| 1612 | f_ret = errno; |
| 1613 | } |
| 1614 | |
| 1615 | if (td->io_ops->close_file) |
| 1616 | ret = td->io_ops->close_file(td, f); |
| 1617 | |
| 1618 | if (!ret) |
| 1619 | ret = f_ret; |
| 1620 | |
| 1621 | td->nr_open_files--; |
| 1622 | fio_file_clear_open(f); |
| 1623 | assert(f->fd == -1); |
| 1624 | return ret; |
| 1625 | } |
| 1626 | |
| 1627 | void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir) |
| 1628 | { |
| 1629 | if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE) |
| 1630 | return; |
| 1631 | |
| 1632 | if (td->o.file_lock_mode == FILE_LOCK_READWRITE) { |
| 1633 | if (ddir == DDIR_READ) |
| 1634 | fio_rwlock_read(f->rwlock); |
| 1635 | else |
| 1636 | fio_rwlock_write(f->rwlock); |
| 1637 | } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) |
| 1638 | fio_mutex_down(f->lock); |
| 1639 | |
| 1640 | td->file_locks[f->fileno] = td->o.file_lock_mode; |
| 1641 | } |
| 1642 | |
| 1643 | void unlock_file(struct thread_data *td, struct fio_file *f) |
| 1644 | { |
| 1645 | if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE) |
| 1646 | return; |
| 1647 | |
| 1648 | if (td->o.file_lock_mode == FILE_LOCK_READWRITE) |
| 1649 | fio_rwlock_unlock(f->rwlock); |
| 1650 | else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) |
| 1651 | fio_mutex_up(f->lock); |
| 1652 | |
| 1653 | td->file_locks[f->fileno] = FILE_LOCK_NONE; |
| 1654 | } |
| 1655 | |
| 1656 | void unlock_file_all(struct thread_data *td, struct fio_file *f) |
| 1657 | { |
| 1658 | if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks) |
| 1659 | return; |
| 1660 | if (td->file_locks[f->fileno] != FILE_LOCK_NONE) |
| 1661 | unlock_file(td, f); |
| 1662 | } |
| 1663 | |
| 1664 | static int recurse_dir(struct thread_data *td, const char *dirname) |
| 1665 | { |
| 1666 | struct dirent *dir; |
| 1667 | int ret = 0; |
| 1668 | DIR *D; |
| 1669 | |
| 1670 | D = opendir(dirname); |
| 1671 | if (!D) { |
| 1672 | char buf[FIO_VERROR_SIZE]; |
| 1673 | |
| 1674 | snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname); |
| 1675 | td_verror(td, errno, buf); |
| 1676 | return 1; |
| 1677 | } |
| 1678 | |
| 1679 | while ((dir = readdir(D)) != NULL) { |
| 1680 | char full_path[PATH_MAX]; |
| 1681 | struct stat sb; |
| 1682 | |
| 1683 | if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, "..")) |
| 1684 | continue; |
| 1685 | |
| 1686 | sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name); |
| 1687 | |
| 1688 | if (lstat(full_path, &sb) == -1) { |
| 1689 | if (errno != ENOENT) { |
| 1690 | td_verror(td, errno, "stat"); |
| 1691 | ret = 1; |
| 1692 | break; |
| 1693 | } |
| 1694 | } |
| 1695 | |
| 1696 | if (S_ISREG(sb.st_mode)) { |
| 1697 | add_file(td, full_path, 0, 1); |
| 1698 | continue; |
| 1699 | } |
| 1700 | if (!S_ISDIR(sb.st_mode)) |
| 1701 | continue; |
| 1702 | |
| 1703 | ret = recurse_dir(td, full_path); |
| 1704 | if (ret) |
| 1705 | break; |
| 1706 | } |
| 1707 | |
| 1708 | closedir(D); |
| 1709 | return ret; |
| 1710 | } |
| 1711 | |
| 1712 | int add_dir_files(struct thread_data *td, const char *path) |
| 1713 | { |
| 1714 | int ret = recurse_dir(td, path); |
| 1715 | |
| 1716 | if (!ret) |
| 1717 | log_info("fio: opendir added %d files\n", td->o.nr_files); |
| 1718 | |
| 1719 | return ret; |
| 1720 | } |
| 1721 | |
| 1722 | void dup_files(struct thread_data *td, struct thread_data *org) |
| 1723 | { |
| 1724 | struct fio_file *f; |
| 1725 | unsigned int i; |
| 1726 | |
| 1727 | dprint(FD_FILE, "dup files: %d\n", org->files_index); |
| 1728 | |
| 1729 | if (!org->files) |
| 1730 | return; |
| 1731 | |
| 1732 | td->files = malloc(org->files_index * sizeof(f)); |
| 1733 | |
| 1734 | if (td->o.file_lock_mode != FILE_LOCK_NONE) |
| 1735 | td->file_locks = malloc(org->files_index); |
| 1736 | |
| 1737 | for_each_file(org, f, i) { |
| 1738 | struct fio_file *__f; |
| 1739 | |
| 1740 | __f = alloc_new_file(td); |
| 1741 | |
| 1742 | if (f->file_name) { |
| 1743 | __f->file_name = smalloc_strdup(f->file_name); |
| 1744 | if (!__f->file_name) |
| 1745 | assert(0); |
| 1746 | |
| 1747 | __f->filetype = f->filetype; |
| 1748 | } |
| 1749 | |
| 1750 | if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) |
| 1751 | __f->lock = f->lock; |
| 1752 | else if (td->o.file_lock_mode == FILE_LOCK_READWRITE) |
| 1753 | __f->rwlock = f->rwlock; |
| 1754 | |
| 1755 | td->files[i] = __f; |
| 1756 | } |
| 1757 | } |
| 1758 | |
| 1759 | /* |
| 1760 | * Returns the index that matches the filename, or -1 if not there |
| 1761 | */ |
| 1762 | int get_fileno(struct thread_data *td, const char *fname) |
| 1763 | { |
| 1764 | struct fio_file *f; |
| 1765 | unsigned int i; |
| 1766 | |
| 1767 | for_each_file(td, f, i) |
| 1768 | if (!strcmp(f->file_name, fname)) |
| 1769 | return i; |
| 1770 | |
| 1771 | return -1; |
| 1772 | } |
| 1773 | |
| 1774 | /* |
| 1775 | * For log usage, where we add/open/close files automatically |
| 1776 | */ |
| 1777 | void free_release_files(struct thread_data *td) |
| 1778 | { |
| 1779 | close_files(td); |
| 1780 | td->o.nr_files = 0; |
| 1781 | td->o.open_files = 0; |
| 1782 | td->files_index = 0; |
| 1783 | td->nr_normal_files = 0; |
| 1784 | } |
| 1785 | |
| 1786 | void fio_file_reset(struct thread_data *td, struct fio_file *f) |
| 1787 | { |
| 1788 | int i; |
| 1789 | |
| 1790 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
| 1791 | f->last_pos[i] = f->file_offset; |
| 1792 | f->last_start[i] = -1ULL; |
| 1793 | } |
| 1794 | |
| 1795 | if (fio_file_axmap(f)) |
| 1796 | axmap_reset(f->io_axmap); |
| 1797 | else if (fio_file_lfsr(f)) |
| 1798 | lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]); |
| 1799 | } |
| 1800 | |
| 1801 | bool fio_files_done(struct thread_data *td) |
| 1802 | { |
| 1803 | struct fio_file *f; |
| 1804 | unsigned int i; |
| 1805 | |
| 1806 | for_each_file(td, f, i) |
| 1807 | if (!fio_file_done(f)) |
| 1808 | return false; |
| 1809 | |
| 1810 | return true; |
| 1811 | } |
| 1812 | |
| 1813 | /* free memory used in initialization phase only */ |
| 1814 | void filesetup_mem_free(void) |
| 1815 | { |
| 1816 | free_already_allocated(); |
| 1817 | } |