filesetup.c

   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 static int native_fallocate(struct thread_data *td, struct fio_file *f)
  42 {
  43         bool success;
  44
  45         success = fio_fallocate(f, 0, f->real_file_size);
  46         dprint(FD_FILE, "native fallocate of file %s size %llu was "
  47                         "%ssuccessful\n", f->file_name,
  48                         (unsigned long long) f->real_file_size,
  49                         !success ? "un": "");
  50
  51         if (success)
  52                 return false;
  53
  54         if (errno == ENOSYS)
  55                 dprint(FD_FILE, "native fallocate is not implemented\n");
  56
  57         return true;
  58 }
  59
  60 static void fallocate_file(struct thread_data *td, struct fio_file *f)
  61 {
  62         int r;
  63
  64         if (td->o.fill_device)
  65                 return;
  66
  67         switch (td->o.fallocate_mode) {
  68         case FIO_FALLOCATE_NATIVE:
  69                 native_fallocate(td, f);
  70                 break;
  71         case FIO_FALLOCATE_NONE:
  72                 break;
  73 #ifdef CONFIG_POSIX_FALLOCATE
  74         case FIO_FALLOCATE_POSIX:
  75                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  76                                  f->file_name,
  77                                  (unsigned long long) f->real_file_size);
  78
  79                 r = posix_fallocate(f->fd, 0, f->real_file_size);
  80                 if (r > 0)
  81                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
  82                 break;
  83 #endif /* CONFIG_POSIX_FALLOCATE */
  84 #ifdef CONFIG_LINUX_FALLOCATE
  85         case FIO_FALLOCATE_KEEP_SIZE:
  86                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
  87                                 "file %s size %llu\n", f->file_name,
  88                                 (unsigned long long) f->real_file_size);
  89
  90                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
  91                 if (r != 0)
  92                         td_verror(td, errno, "fallocate");
  93
  94                 break;
  95 #endif /* CONFIG_LINUX_FALLOCATE */
  96         default:
  97                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
  98                 assert(0);
  99         }
 100 }
 101
 102 /*
 103  * Leaves f->fd open on success, caller must close
 104  */
 105 static int extend_file(struct thread_data *td, struct fio_file *f)
 106 {
 107         int new_layout = 0, unlink_file = 0, flags;
 108         unsigned long long left;
 109         unsigned int bs;
 110         char *b = NULL;
 111
 112         if (read_only) {
 113                 log_err("fio: refusing extend of file due to read-only\n");
 114                 return 0;
 115         }
 116
 117         /*
 118          * check if we need to lay the file out complete again. fio
 119          * does that for operations involving reads, or for writes
 120          * where overwrite is set
 121          */
 122         if (td_read(td) ||
 123            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
 124             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
 125                 new_layout = 1;
 126         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
 127                 unlink_file = 1;
 128
 129         if (unlink_file || new_layout) {
 130                 int ret;
 131
 132                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
 133
 134                 ret = td_io_unlink_file(td, f);
 135                 if (ret != 0 && ret != ENOENT) {
 136                         td_verror(td, errno, "unlink");
 137                         return 1;
 138                 }
 139         }
 140
 141         flags = O_WRONLY;
 142         if (td->o.allow_create)
 143                 flags |= O_CREAT;
 144         if (new_layout)
 145                 flags |= O_TRUNC;
 146
 147 #ifdef WIN32
 148         flags |= _O_BINARY;
 149 #endif
 150
 151         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 152         f->fd = open(f->file_name, flags, 0644);
 153         if (f->fd < 0) {
 154                 int err = errno;
 155
 156                 if (err == ENOENT && !td->o.allow_create)
 157                         log_err("fio: file creation disallowed by "
 158                                         "allow_file_create=0\n");
 159                 else
 160                         td_verror(td, err, "open");
 161                 return 1;
 162         }
 163
 164         fallocate_file(td, f);
 165
 166         /*
 167          * If our jobs don't require regular files initially, we're done.
 168          */
 169         if (!new_layout)
 170                 goto done;
 171
 172         /*
 173          * The size will be -1ULL when fill_device is used, so don't truncate
 174          * or fallocate this file, just write it
 175          */
 176         if (!td->o.fill_device) {
 177                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 178                                         (unsigned long long) f->real_file_size);
 179                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 180                         if (errno != EFBIG) {
 181                                 td_verror(td, errno, "ftruncate");
 182                                 goto err;
 183                         }
 184                 }
 185         }
 186
 187         left = f->real_file_size;
 188         bs = td->o.max_bs[DDIR_WRITE];
 189         if (bs > left)
 190                 bs = left;
 191
 192         b = malloc(bs);
 193         if (!b) {
 194                 td_verror(td, errno, "malloc");
 195                 goto err;
 196         }
 197
 198         while (left && !td->terminate) {
 199                 ssize_t r;
 200
 201                 if (bs > left)
 202                         bs = left;
 203
 204                 fill_io_buffer(td, b, bs, bs);
 205
 206                 r = write(f->fd, b, bs);
 207
 208                 if (r > 0) {
 209                         left -= r;
 210                         continue;
 211                 } else {
 212                         if (r < 0) {
 213                                 int __e = errno;
 214
 215                                 if (__e == ENOSPC) {
 216                                         if (td->o.fill_device)
 217                                                 break;
 218                                         log_info("fio: ENOSPC on laying out "
 219                                                  "file, stopping\n");
 220                                         break;
 221                                 }
 222                                 td_verror(td, errno, "write");
 223                         } else
 224                                 td_verror(td, EIO, "write");
 225
 226                         break;
 227                 }
 228         }
 229
 230         if (td->terminate) {
 231                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 232                 td_io_unlink_file(td, f);
 233         } else if (td->o.create_fsync) {
 234                 if (fsync(f->fd) < 0) {
 235                         td_verror(td, errno, "fsync");
 236                         goto err;
 237                 }
 238         }
 239         if (td->o.fill_device && !td_write(td)) {
 240                 fio_file_clear_size_known(f);
 241                 if (td_io_get_file_size(td, f))
 242                         goto err;
 243                 if (f->io_size > f->real_file_size)
 244                         f->io_size = f->real_file_size;
 245         }
 246
 247         free(b);
 248 done:
 249         return 0;
 250 err:
 251         close(f->fd);
 252         f->fd = -1;
 253         if (b)
 254                 free(b);
 255         return 1;
 256 }
 257
 258 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 259 {
 260         int ret = 0, r, did_open = 0, old_runstate;
 261         unsigned long long left;
 262         unsigned int bs;
 263         char *b;
 264
 265         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 266             td_ioengine_flagged(td, FIO_NOIO))
 267                 return 0;
 268
 269         if (f->filetype == FIO_TYPE_CHAR)
 270                 return 0;
 271
 272         if (!fio_file_open(f)) {
 273                 if (td->io_ops->open_file(td, f)) {
 274                         log_err("fio: cannot pre-read, failed to open file\n");
 275                         return 1;
 276                 }
 277                 did_open = 1;
 278         }
 279
 280         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 281
 282         left = f->io_size;
 283         bs = td->o.max_bs[DDIR_READ];
 284         if (bs > left)
 285                 bs = left;
 286
 287         b = malloc(bs);
 288         if (!b) {
 289                 td_verror(td, errno, "malloc");
 290                 ret = 1;
 291                 goto error;
 292         }
 293         memset(b, 0, bs);
 294
 295         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 296                 td_verror(td, errno, "lseek");
 297                 log_err("fio: failed to lseek pre-read file\n");
 298                 ret = 1;
 299                 goto error;
 300         }
 301
 302         while (left && !td->terminate) {
 303                 if (bs > left)
 304                         bs = left;
 305
 306                 r = read(f->fd, b, bs);
 307
 308                 if (r == (int) bs) {
 309                         left -= bs;
 310                         continue;
 311                 } else {
 312                         td_verror(td, EIO, "pre_read");
 313                         break;
 314                 }
 315         }
 316
 317 error:
 318         td_restore_runstate(td, old_runstate);
 319
 320         if (did_open)
 321                 td->io_ops->close_file(td, f);
 322
 323         free(b);
 324         return ret;
 325 }
 326
 327 unsigned long long get_rand_file_size(struct thread_data *td)
 328 {
 329         unsigned long long ret, sized;
 330         uint64_t frand_max;
 331         unsigned long r;
 332
 333         frand_max = rand_max(&td->file_size_state);
 334         r = __rand(&td->file_size_state);
 335         sized = td->o.file_size_high - td->o.file_size_low;
 336         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 337         ret += td->o.file_size_low;
 338         ret -= (ret % td->o.rw_min_bs);
 339         return ret;
 340 }
 341
 342 static int file_size(struct thread_data *td, struct fio_file *f)
 343 {
 344         struct stat st;
 345
 346         if (stat(f->file_name, &st) == -1) {
 347                 td_verror(td, errno, "fstat");
 348                 return 1;
 349         }
 350
 351         f->real_file_size = st.st_size;
 352         return 0;
 353 }
 354
 355 static int bdev_size(struct thread_data *td, struct fio_file *f)
 356 {
 357         unsigned long long bytes = 0;
 358         int r;
 359
 360         if (td->io_ops->open_file(td, f)) {
 361                 log_err("fio: failed opening blockdev %s for size check\n",
 362                         f->file_name);
 363                 return 1;
 364         }
 365
 366         r = blockdev_size(f, &bytes);
 367         if (r) {
 368                 td_verror(td, r, "blockdev_size");
 369                 goto err;
 370         }
 371
 372         if (!bytes) {
 373                 log_err("%s: zero sized block device?\n", f->file_name);
 374                 goto err;
 375         }
 376
 377         f->real_file_size = bytes;
 378         td->io_ops->close_file(td, f);
 379         return 0;
 380 err:
 381         td->io_ops->close_file(td, f);
 382         return 1;
 383 }
 384
 385 static int char_size(struct thread_data *td, struct fio_file *f)
 386 {
 387 #ifdef FIO_HAVE_CHARDEV_SIZE
 388         unsigned long long bytes = 0;
 389         int r;
 390
 391         if (td->io_ops->open_file(td, f)) {
 392                 log_err("fio: failed opening chardev %s for size check\n",
 393                         f->file_name);
 394                 return 1;
 395         }
 396
 397         r = chardev_size(f, &bytes);
 398         if (r) {
 399                 td_verror(td, r, "chardev_size");
 400                 goto err;
 401         }
 402
 403         if (!bytes) {
 404                 log_err("%s: zero sized char device?\n", f->file_name);
 405                 goto err;
 406         }
 407
 408         f->real_file_size = bytes;
 409         td->io_ops->close_file(td, f);
 410         return 0;
 411 err:
 412         td->io_ops->close_file(td, f);
 413         return 1;
 414 #else
 415         f->real_file_size = -1ULL;
 416         return 0;
 417 #endif
 418 }
 419
 420 static int get_file_size(struct thread_data *td, struct fio_file *f)
 421 {
 422         int ret = 0;
 423
 424         if (fio_file_size_known(f))
 425                 return 0;
 426
 427         if (f->filetype == FIO_TYPE_FILE)
 428                 ret = file_size(td, f);
 429         else if (f->filetype == FIO_TYPE_BLOCK)
 430                 ret = bdev_size(td, f);
 431         else if (f->filetype == FIO_TYPE_CHAR)
 432                 ret = char_size(td, f);
 433         else
 434                 f->real_file_size = -1ULL;
 435
 436         /*
 437          * Leave ->real_file_size with 0 since it could be expectation
 438          * of initial setup for regular files.
 439          */
 440         if (ret)
 441                 return ret;
 442
 443         /*
 444          * If ->real_file_size is -1, a conditional for the message
 445          * "offset extends end" is always true, but it makes no sense,
 446          * so just return the same value here.
 447          */
 448         if (f->real_file_size == -1ULL) {
 449                 log_info("%s: failed to get file size of %s\n", td->o.name,
 450                                         f->file_name);
 451                 return 1;
 452         }
 453
 454         if (td->o.start_offset && f->file_offset == 0)
 455                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 456                                         f->file_name);
 457         /*
 458          * ->file_offset normally hasn't been initialized yet, so this
 459          * is basically always false.
 460          */
 461         if (f->file_offset > f->real_file_size) {
 462                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 463                                         (unsigned long long) f->file_offset,
 464                                         (unsigned long long) f->real_file_size);
 465                 return 1;
 466         }
 467
 468         fio_file_set_size_known(f);
 469         return 0;
 470 }
 471
 472 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 473                                    unsigned long long off,
 474                                    unsigned long long len)
 475 {
 476         int errval = 0, ret = 0;
 477
 478 #ifdef CONFIG_ESX
 479         return 0;
 480 #endif
 481
 482         if (len == -1ULL)
 483                 len = f->io_size;
 484         if (off == -1ULL)
 485                 off = f->file_offset;
 486
 487         if (len == -1ULL || off == -1ULL)
 488                 return 0;
 489
 490         if (td->io_ops->invalidate) {
 491                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 492                         f->file_name);
 493                 ret = td->io_ops->invalidate(td, f);
 494                 if (ret < 0)
 495                         errval = -ret;
 496         } else if (f->filetype == FIO_TYPE_FILE) {
 497                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 498                         f->file_name, off, len);
 499                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 500                 if (ret)
 501                         errval = ret;
 502         } else if (f->filetype == FIO_TYPE_BLOCK) {
 503                 int retry_count = 0;
 504
 505                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 506                 ret = blockdev_invalidate_cache(f);
 507                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 508                         /*
 509                          * Linux multipath devices reject ioctl while
 510                          * the maps are being updated. That window can
 511                          * last tens of milliseconds; we'll try up to
 512                          * a quarter of a second.
 513                          */
 514                         usleep(10000);
 515                         ret = blockdev_invalidate_cache(f);
 516                 }
 517                 if (ret < 0 && errno == EACCES && geteuid()) {
 518                         if (!root_warn) {
 519                                 log_err("fio: only root may flush block "
 520                                         "devices. Cache flush bypassed!\n");
 521                                 root_warn = 1;
 522                         }
 523                         ret = 0;
 524                 }
 525                 if (ret < 0)
 526                         errval = errno;
 527         } else if (f->filetype == FIO_TYPE_CHAR ||
 528                    f->filetype == FIO_TYPE_PIPE) {
 529                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 530                 ret = 0;
 531         }
 532
 533         /*
 534          * Cache flushing isn't a fatal condition, and we know it will
 535          * happen on some platforms where we don't have the proper
 536          * function to flush eg block device caches. So just warn and
 537          * continue on our way.
 538          */
 539         if (errval)
 540                 log_info("fio: cache invalidation of %s failed: %s\n",
 541                          f->file_name, strerror(errval));
 542
 543         return 0;
 544
 545 }
 546
 547 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 548 {
 549         if (!fio_file_open(f))
 550                 return 0;
 551
 552         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 553 }
 554
 555 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 556 {
 557         int ret = 0;
 558
 559         dprint(FD_FILE, "fd close %s\n", f->file_name);
 560
 561         remove_file_hash(f);
 562
 563         if (close(f->fd) < 0)
 564                 ret = errno;
 565
 566         f->fd = -1;
 567
 568         if (f->shadow_fd != -1) {
 569                 close(f->shadow_fd);
 570                 f->shadow_fd = -1;
 571         }
 572
 573         f->engine_pos = 0;
 574         return ret;
 575 }
 576
 577 int file_lookup_open(struct fio_file *f, int flags)
 578 {
 579         struct fio_file *__f;
 580         int from_hash;
 581
 582         __f = lookup_file_hash(f->file_name);
 583         if (__f) {
 584                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 585                 f->lock = __f->lock;
 586                 from_hash = 1;
 587         } else {
 588                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 589                 from_hash = 0;
 590         }
 591
 592 #ifdef WIN32
 593         flags |= _O_BINARY;
 594 #endif
 595
 596         f->fd = open(f->file_name, flags, 0600);
 597         return from_hash;
 598 }
 599
 600 static int file_close_shadow_fds(struct thread_data *td)
 601 {
 602         struct fio_file *f;
 603         int num_closed = 0;
 604         unsigned int i;
 605
 606         for_each_file(td, f, i) {
 607                 if (f->shadow_fd == -1)
 608                         continue;
 609
 610                 close(f->shadow_fd);
 611                 f->shadow_fd = -1;
 612                 num_closed++;
 613         }
 614
 615         return num_closed;
 616 }
 617
 618 int generic_open_file(struct thread_data *td, struct fio_file *f)
 619 {
 620         int is_std = 0;
 621         int flags = 0;
 622         int from_hash = 0;
 623
 624         dprint(FD_FILE, "fd open %s\n", f->file_name);
 625
 626         if (!strcmp(f->file_name, "-")) {
 627                 if (td_rw(td)) {
 628                         log_err("fio: can't read/write to stdin/out\n");
 629                         return 1;
 630                 }
 631                 is_std = 1;
 632
 633                 /*
 634                  * move output logging to stderr, if we are writing to stdout
 635                  */
 636                 if (td_write(td))
 637                         f_out = stderr;
 638         }
 639
 640         if (td_trim(td))
 641                 goto skip_flags;
 642         if (td->o.odirect)
 643                 flags |= OS_O_DIRECT;
 644         if (td->o.oatomic) {
 645                 if (!FIO_O_ATOMIC) {
 646                         td_verror(td, EINVAL, "OS does not support atomic IO");
 647                         return 1;
 648                 }
 649                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 650         }
 651         if (td->o.sync_io)
 652                 flags |= O_SYNC;
 653         if (td->o.create_on_open && td->o.allow_create)
 654                 flags |= O_CREAT;
 655 skip_flags:
 656         if (f->filetype != FIO_TYPE_FILE)
 657                 flags |= FIO_O_NOATIME;
 658
 659 open_again:
 660         if (td_write(td)) {
 661                 if (!read_only)
 662                         flags |= O_RDWR;
 663
 664                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 665                         flags |= O_CREAT;
 666
 667                 if (is_std)
 668                         f->fd = dup(STDOUT_FILENO);
 669                 else
 670                         from_hash = file_lookup_open(f, flags);
 671         } else if (td_read(td)) {
 672                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 673                         flags |= O_RDWR;
 674                 else
 675                         flags |= O_RDONLY;
 676
 677                 if (is_std)
 678                         f->fd = dup(STDIN_FILENO);
 679                 else
 680                         from_hash = file_lookup_open(f, flags);
 681         } else if (td_trim(td)) {
 682                 assert(!td_rw(td)); /* should have matched above */
 683                 flags |= O_RDWR;
 684                 from_hash = file_lookup_open(f, flags);
 685         }
 686
 687         if (f->fd == -1) {
 688                 char buf[FIO_VERROR_SIZE];
 689                 int __e = errno;
 690
 691                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 692                         flags &= ~FIO_O_NOATIME;
 693                         goto open_again;
 694                 }
 695                 if (__e == EMFILE && file_close_shadow_fds(td))
 696                         goto open_again;
 697
 698                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 699
 700                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 701                         log_err("fio: looks like your file system does not " \
 702                                 "support direct=1/buffered=0\n");
 703                 }
 704
 705                 td_verror(td, __e, buf);
 706                 return 1;
 707         }
 708
 709         if (!from_hash && f->fd != -1) {
 710                 if (add_file_hash(f)) {
 711                         int fio_unused ret;
 712
 713                         /*
 714                          * Stash away descriptor for later close. This is to
 715                          * work-around a "feature" on Linux, where a close of
 716                          * an fd that has been opened for write will trigger
 717                          * udev to call blkid to check partitions, fs id, etc.
 718                          * That pollutes the device cache, which can slow down
 719                          * unbuffered accesses.
 720                          */
 721                         if (f->shadow_fd == -1)
 722                                 f->shadow_fd = f->fd;
 723                         else {
 724                                 /*
 725                                  * OK to ignore, we haven't done anything
 726                                  * with it
 727                                  */
 728                                 ret = generic_close_file(td, f);
 729                         }
 730                         goto open_again;
 731                 }
 732         }
 733
 734         return 0;
 735 }
 736
 737 /*
 738  * This function i.e. get_file_size() is the default .get_file_size
 739  * implementation of majority of I/O engines.
 740  */
 741 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 742 {
 743         return get_file_size(td, f);
 744 }
 745
 746 /*
 747  * open/close all files, so that ->real_file_size gets set
 748  */
 749 static int get_file_sizes(struct thread_data *td)
 750 {
 751         struct fio_file *f;
 752         unsigned int i;
 753         int err = 0;
 754
 755         for_each_file(td, f, i) {
 756                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 757                                                                 f->file_name);
 758
 759                 if (td_io_get_file_size(td, f)) {
 760                         if (td->error != ENOENT) {
 761                                 log_err("%s\n", td->verror);
 762                                 err = 1;
 763                                 break;
 764                         }
 765                         clear_error(td);
 766                 }
 767
 768                 /*
 769                  * There are corner cases where we end up with -1 for
 770                  * ->real_file_size due to unsupported file type, etc.
 771                  * We then just set to size option value divided by number
 772                  * of files, similar to the way file ->io_size is set.
 773                  * stat(2) failure doesn't set ->real_file_size to -1.
 774                  */
 775                 if (f->real_file_size == -1ULL && td->o.size)
 776                         f->real_file_size = td->o.size / td->o.nr_files;
 777         }
 778
 779         return err;
 780 }
 781
 782 struct fio_mount {
 783         struct flist_head list;
 784         const char *base;
 785         char __base[256];
 786         unsigned int key;
 787 };
 788
 789 /*
 790  * Get free number of bytes for each file on each unique mount.
 791  */
 792 static unsigned long long get_fs_free_counts(struct thread_data *td)
 793 {
 794         struct flist_head *n, *tmp;
 795         unsigned long long ret = 0;
 796         struct fio_mount *fm;
 797         FLIST_HEAD(list);
 798         struct fio_file *f;
 799         unsigned int i;
 800
 801         for_each_file(td, f, i) {
 802                 struct stat sb;
 803                 char buf[256];
 804
 805                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 806                         if (f->real_file_size != -1ULL)
 807                                 ret += f->real_file_size;
 808                         continue;
 809                 } else if (f->filetype != FIO_TYPE_FILE)
 810                         continue;
 811
 812                 buf[255] = '\0';
 813                 strncpy(buf, f->file_name, 255);
 814
 815                 if (stat(buf, &sb) < 0) {
 816                         if (errno != ENOENT)
 817                                 break;
 818                         strcpy(buf, ".");
 819                         if (stat(buf, &sb) < 0)
 820                                 break;
 821                 }
 822
 823                 fm = NULL;
 824                 flist_for_each(n, &list) {
 825                         fm = flist_entry(n, struct fio_mount, list);
 826                         if (fm->key == sb.st_dev)
 827                                 break;
 828
 829                         fm = NULL;
 830                 }
 831
 832                 if (fm)
 833                         continue;
 834
 835                 fm = calloc(1, sizeof(*fm));
 836                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 837                 fm->base = basename(fm->__base);
 838                 fm->key = sb.st_dev;
 839                 flist_add(&fm->list, &list);
 840         }
 841
 842         flist_for_each_safe(n, tmp, &list) {
 843                 unsigned long long sz;
 844
 845                 fm = flist_entry(n, struct fio_mount, list);
 846                 flist_del(&fm->list);
 847
 848                 sz = get_fs_free_size(fm->base);
 849                 if (sz && sz != -1ULL)
 850                         ret += sz;
 851
 852                 free(fm);
 853         }
 854
 855         return ret;
 856 }
 857
 858 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 859 {
 860         struct thread_options *o = &td->o;
 861         unsigned long long align_bs;
 862         unsigned long long offset;
 863
 864         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 865                 return f->real_file_size;
 866
 867         if (o->start_offset_percent > 0) {
 868                 /*
 869                  * if offset_align is provided, set initial offset
 870                  */
 871                 if (fio_option_is_set(o, start_offset_align)) {
 872                         align_bs = o->start_offset_align;
 873                 } else {
 874                         /* else take the minimum block size */
 875                         align_bs = td_min_bs(td);
 876                 }
 877
 878                 /* calculate the raw offset */
 879                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 880                         (td->subjob_number * o->offset_increment);
 881
 882                 /*
 883                  * block align the offset at the next available boundary at
 884                  * ceiling(offset / align_bs) * align_bs
 885                  */
 886                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 887
 888         } else {
 889                 /* start_offset_percent not set */
 890                 offset = o->start_offset +
 891                                 td->subjob_number * o->offset_increment;
 892         }
 893
 894         return offset;
 895 }
 896
 897 /*
 898  * Open the files and setup files sizes, creating files if necessary.
 899  */
 900 int setup_files(struct thread_data *td)
 901 {
 902         unsigned long long total_size, extend_size;
 903         struct thread_options *o = &td->o;
 904         struct fio_file *f;
 905         unsigned int i, nr_fs_extra = 0;
 906         int err = 0, need_extend;
 907         int old_state;
 908         const unsigned int bs = td_min_bs(td);
 909         uint64_t fs = 0;
 910
 911         dprint(FD_FILE, "setup files\n");
 912
 913         old_state = td_bump_runstate(td, TD_SETTING_UP);
 914
 915         if (o->read_iolog_file)
 916                 goto done;
 917
 918         /*
 919          * Find out physical size of files or devices for this thread,
 920          * before we determine I/O size and range of our targets.
 921          * If ioengine defines a setup() method, it's responsible for
 922          * opening the files and setting f->real_file_size to indicate
 923          * the valid range for that file.
 924          */
 925         if (td->io_ops->setup)
 926                 err = td->io_ops->setup(td);
 927         else
 928                 err = get_file_sizes(td);
 929
 930         if (err)
 931                 goto err_out;
 932
 933         /*
 934          * check sizes. if the files/devices do not exist and the size
 935          * isn't passed to fio, abort.
 936          */
 937         total_size = 0;
 938         for_each_file(td, f, i) {
 939                 f->fileno = i;
 940                 if (f->real_file_size == -1ULL)
 941                         total_size = -1ULL;
 942                 else
 943                         total_size += f->real_file_size;
 944         }
 945
 946         if (o->fill_device)
 947                 td->fill_device_size = get_fs_free_counts(td);
 948
 949         /*
 950          * device/file sizes are zero and no size given, punt
 951          */
 952         if ((!total_size || total_size == -1ULL) && !o->size &&
 953             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 954             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 955                 log_err("%s: you need to specify size=\n", o->name);
 956                 td_verror(td, EINVAL, "total_file_size");
 957                 goto err_out;
 958         }
 959
 960         /*
 961          * Calculate per-file size and potential extra size for the
 962          * first files, if needed (i.e. if we don't have a fixed size).
 963          */
 964         if (!o->file_size_low && o->nr_files) {
 965                 uint64_t all_fs;
 966
 967                 fs = o->size / o->nr_files;
 968                 all_fs = fs * o->nr_files;
 969
 970                 if (all_fs < o->size)
 971                         nr_fs_extra = (o->size - all_fs) / bs;
 972         }
 973
 974         /*
 975          * now file sizes are known, so we can set ->io_size. if size= is
 976          * not given, ->io_size is just equal to ->real_file_size. if size
 977          * is given, ->io_size is size / nr_files.
 978          */
 979         extend_size = total_size = 0;
 980         need_extend = 0;
 981         for_each_file(td, f, i) {
 982                 f->file_offset = get_start_offset(td, f);
 983
 984                 /*
 985                  * Update ->io_size depending on options specified.
 986                  * ->file_size_low being 0 means filesize option isn't set.
 987                  * Non zero ->file_size_low equals ->file_size_high means
 988                  * filesize option is set in a fixed size format.
 989                  * Non zero ->file_size_low not equals ->file_size_high means
 990                  * filesize option is set in a range format.
 991                  */
 992                 if (!o->file_size_low) {
 993                         /*
 994                          * no file size or range given, file size is equal to
 995                          * total size divided by number of files. If the size
 996                          * doesn't divide nicely with the min blocksize,
 997                          * make the first files bigger.
 998                          */
 999                         f->io_size = fs;
1000                         if (nr_fs_extra) {
1001                                 nr_fs_extra--;
1002                                 f->io_size += bs;
1003                         }
1004
1005                         /*
1006                          * We normally don't come here for regular files, but
1007                          * if the result is 0 for a regular file, set it to the
1008                          * real file size. This could be size of the existing
1009                          * one if it already exists, but otherwise will be set
1010                          * to 0. A new file won't be created because
1011                          * ->io_size + ->file_offset equals ->real_file_size.
1012                          */
1013                         if (!f->io_size) {
1014                                 if (f->file_offset > f->real_file_size)
1015                                         goto err_offset;
1016                                 f->io_size = f->real_file_size - f->file_offset;
1017                                 if (!f->io_size)
1018                                         log_info("fio: file %s may be ignored\n",
1019                                                 f->file_name);
1020                         }
1021                 } else if (f->real_file_size < o->file_size_low ||
1022                            f->real_file_size > o->file_size_high) {
1023                         if (f->file_offset > o->file_size_low)
1024                                 goto err_offset;
1025                         /*
1026                          * file size given. if it's fixed, use that. if it's a
1027                          * range, generate a random size in-between.
1028                          */
1029                         if (o->file_size_low == o->file_size_high)
1030                                 f->io_size = o->file_size_low - f->file_offset;
1031                         else {
1032                                 f->io_size = get_rand_file_size(td)
1033                                                 - f->file_offset;
1034                         }
1035                 } else
1036                         f->io_size = f->real_file_size - f->file_offset;
1037
1038                 if (f->io_size == -1ULL)
1039                         total_size = -1ULL;
1040                 else {
1041                         if (o->size_percent) {
1042                                 uint64_t file_size;
1043
1044                                 file_size = f->io_size + f->file_offset;
1045                                 f->io_size = (file_size *
1046                                               o->size_percent) / 100;
1047                                 if (f->io_size > (file_size - f->file_offset))
1048                                         f->io_size = file_size - f->file_offset;
1049
1050                                 f->io_size -= (f->io_size % td_min_bs(td));
1051                         }
1052                         total_size += f->io_size;
1053                 }
1054
1055                 if (f->filetype == FIO_TYPE_FILE &&
1056                     (f->io_size + f->file_offset) > f->real_file_size &&
1057                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1058                         if (!o->create_on_open) {
1059                                 need_extend++;
1060                                 extend_size += (f->io_size + f->file_offset);
1061                                 fio_file_set_extend(f);
1062                         } else
1063                                 f->real_file_size = f->io_size + f->file_offset;
1064                 }
1065         }
1066
1067         if (td->o.block_error_hist) {
1068                 int len;
1069
1070                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1071                 f = td->files[0];
1072                 len = f->io_size / td->o.bs[DDIR_TRIM];
1073                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1074                         log_err("fio: cannot calculate block histogram with "
1075                                 "%d trim blocks, maximum %d\n",
1076                                 len, MAX_NR_BLOCK_INFOS);
1077                         td_verror(td, EINVAL, "block_error_hist");
1078                         goto err_out;
1079                 }
1080
1081                 td->ts.nr_block_infos = len;
1082                 for (i = 0; i < len; i++)
1083                         td->ts.block_infos[i] =
1084                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1085         } else
1086                 td->ts.nr_block_infos = 0;
1087
1088         if (!o->size || (total_size && o->size > total_size))
1089                 o->size = total_size;
1090
1091         if (o->size < td_min_bs(td)) {
1092                 log_err("fio: blocksize too large for data set\n");
1093                 goto err_out;
1094         }
1095
1096         /*
1097          * See if we need to extend some files, typically needed when our
1098          * target regular files don't exist yet, but our jobs require them
1099          * initially due to read I/Os.
1100          */
1101         if (need_extend) {
1102                 temp_stall_ts = 1;
1103                 if (output_format & FIO_OUTPUT_NORMAL) {
1104                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1105                                  o->name,
1106                                  need_extend > 1 ? "s" : "",
1107                                  need_extend,
1108                                  need_extend > 1 ? "s" : "",
1109                                  need_extend > 1 ? "total " : "",
1110                                  extend_size >> 20);
1111                 }
1112
1113                 for_each_file(td, f, i) {
1114                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1115
1116                         if (!fio_file_extend(f))
1117                                 continue;
1118
1119                         assert(f->filetype == FIO_TYPE_FILE);
1120                         fio_file_clear_extend(f);
1121                         if (!o->fill_device) {
1122                                 old_len = f->real_file_size;
1123                                 extend_len = f->io_size + f->file_offset -
1124                                                 old_len;
1125                         }
1126                         f->real_file_size = (f->io_size + f->file_offset);
1127                         err = extend_file(td, f);
1128                         if (err)
1129                                 break;
1130
1131                         err = __file_invalidate_cache(td, f, old_len,
1132                                                                 extend_len);
1133
1134                         /*
1135                          * Shut up static checker
1136                          */
1137                         if (f->fd != -1)
1138                                 close(f->fd);
1139
1140                         f->fd = -1;
1141                         if (err)
1142                                 break;
1143                 }
1144                 temp_stall_ts = 0;
1145         }
1146
1147         if (err)
1148                 goto err_out;
1149
1150         if (!o->zone_size)
1151                 o->zone_size = o->size;
1152
1153         /*
1154          * iolog already set the total io size, if we read back
1155          * stored entries.
1156          */
1157         if (!o->read_iolog_file) {
1158                 if (o->io_size)
1159                         td->total_io_size = o->io_size * o->loops;
1160                 else
1161                         td->total_io_size = o->size * o->loops;
1162         }
1163
1164 done:
1165         if (o->create_only)
1166                 td->done = 1;
1167
1168         td_restore_runstate(td, old_state);
1169         return 0;
1170 err_offset:
1171         log_err("%s: you need to specify valid offset=\n", o->name);
1172 err_out:
1173         td_restore_runstate(td, old_state);
1174         return 1;
1175 }
1176
1177 int pre_read_files(struct thread_data *td)
1178 {
1179         struct fio_file *f;
1180         unsigned int i;
1181
1182         dprint(FD_FILE, "pre_read files\n");
1183
1184         for_each_file(td, f, i) {
1185                 if (pre_read_file(td, f))
1186                         return -1;
1187         }
1188
1189         return 0;
1190 }
1191
1192 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1193 {
1194         unsigned int range_size, seed;
1195         unsigned long nranges;
1196         uint64_t fsize;
1197
1198         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1199         fsize = min(f->real_file_size, f->io_size);
1200
1201         nranges = (fsize + range_size - 1) / range_size;
1202
1203         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1204         if (!td->o.rand_repeatable)
1205                 seed = td->rand_seeds[4];
1206
1207         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1208                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1209         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1210                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1211         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1212                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1213
1214         return 1;
1215 }
1216
1217 static int init_rand_distribution(struct thread_data *td)
1218 {
1219         struct fio_file *f;
1220         unsigned int i;
1221         int state;
1222
1223         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1224                 return 0;
1225
1226         state = td_bump_runstate(td, TD_SETTING_UP);
1227
1228         for_each_file(td, f, i)
1229                 __init_rand_distribution(td, f);
1230
1231         td_restore_runstate(td, state);
1232
1233         return 1;
1234 }
1235
1236 /*
1237  * Check if the number of blocks exceeds the randomness capability of
1238  * the selected generator. Tausworthe is 32-bit, the others are fullly
1239  * 64-bit capable.
1240  */
1241 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1242                                  uint64_t blocks)
1243 {
1244         if (blocks <= FRAND32_MAX)
1245                 return 0;
1246         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1247                 return 0;
1248
1249         /*
1250          * If the user hasn't specified a random generator, switch
1251          * to tausworthe64 with informational warning. If the user did
1252          * specify one, just warn.
1253          */
1254         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1255                         f->file_name);
1256
1257         if (!fio_option_is_set(&td->o, random_generator)) {
1258                 log_info("fio: Switching to tausworthe64. Use the "
1259                          "random_generator= option to get rid of this "
1260                          "warning.\n");
1261                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1262                 return 0;
1263         }
1264
1265         /*
1266          * Just make this information to avoid breaking scripts.
1267          */
1268         log_info("fio: Use the random_generator= option to switch to lfsr or "
1269                          "tausworthe64.\n");
1270         return 0;
1271 }
1272
1273 int init_random_map(struct thread_data *td)
1274 {
1275         unsigned long long blocks;
1276         struct fio_file *f;
1277         unsigned int i;
1278
1279         if (init_rand_distribution(td))
1280                 return 0;
1281         if (!td_random(td))
1282                 return 0;
1283
1284         for_each_file(td, f, i) {
1285                 uint64_t fsize = min(f->real_file_size, f->io_size);
1286
1287                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1288
1289                 if (check_rand_gen_limits(td, f, blocks))
1290                         return 1;
1291
1292                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1293                         unsigned long seed;
1294
1295                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1296
1297                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1298                                 fio_file_set_lfsr(f);
1299                                 continue;
1300                         }
1301                 } else if (!td->o.norandommap) {
1302                         f->io_axmap = axmap_new(blocks);
1303                         if (f->io_axmap) {
1304                                 fio_file_set_axmap(f);
1305                                 continue;
1306                         }
1307                 } else if (td->o.norandommap)
1308                         continue;
1309
1310                 if (!td->o.softrandommap) {
1311                         log_err("fio: failed allocating random map. If running"
1312                                 " a large number of jobs, try the 'norandommap'"
1313                                 " option or set 'softrandommap'. Or give"
1314                                 " a larger --alloc-size to fio.\n");
1315                         return 1;
1316                 }
1317
1318                 log_info("fio: file %s failed allocating random map. Running "
1319                          "job without.\n", f->file_name);
1320         }
1321
1322         return 0;
1323 }
1324
1325 void close_files(struct thread_data *td)
1326 {
1327         struct fio_file *f;
1328         unsigned int i;
1329
1330         for_each_file(td, f, i) {
1331                 if (fio_file_open(f))
1332                         td_io_close_file(td, f);
1333         }
1334 }
1335
1336 void close_and_free_files(struct thread_data *td)
1337 {
1338         struct fio_file *f;
1339         unsigned int i;
1340         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1341
1342         dprint(FD_FILE, "close files\n");
1343
1344         for_each_file(td, f, i) {
1345                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1346                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1347                         td_io_unlink_file(td, f);
1348                 }
1349
1350                 if (fio_file_open(f))
1351                         td_io_close_file(td, f);
1352
1353                 remove_file_hash(f);
1354
1355                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1356                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1357                         td_io_unlink_file(td, f);
1358                 }
1359
1360                 if (use_free)
1361                         free(f->file_name);
1362                 else
1363                         sfree(f->file_name);
1364                 f->file_name = NULL;
1365                 if (fio_file_axmap(f)) {
1366                         axmap_free(f->io_axmap);
1367                         f->io_axmap = NULL;
1368                 }
1369                 if (use_free)
1370                         free(f);
1371                 else
1372                         sfree(f);
1373         }
1374
1375         td->o.filename = NULL;
1376         free(td->files);
1377         free(td->file_locks);
1378         td->files_index = 0;
1379         td->files = NULL;
1380         td->file_locks = NULL;
1381         td->o.file_lock_mode = FILE_LOCK_NONE;
1382         td->o.nr_files = 0;
1383 }
1384
1385 static void get_file_type(struct fio_file *f)
1386 {
1387         struct stat sb;
1388
1389         if (!strcmp(f->file_name, "-"))
1390                 f->filetype = FIO_TYPE_PIPE;
1391         else
1392                 f->filetype = FIO_TYPE_FILE;
1393
1394 #ifdef WIN32
1395         /* \\.\ is the device namespace in Windows, where every file is
1396          * a block device */
1397         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1398                 f->filetype = FIO_TYPE_BLOCK;
1399 #endif
1400
1401         if (!stat(f->file_name, &sb)) {
1402                 if (S_ISBLK(sb.st_mode))
1403                         f->filetype = FIO_TYPE_BLOCK;
1404                 else if (S_ISCHR(sb.st_mode))
1405                         f->filetype = FIO_TYPE_CHAR;
1406                 else if (S_ISFIFO(sb.st_mode))
1407                         f->filetype = FIO_TYPE_PIPE;
1408         }
1409 }
1410
1411 static bool __is_already_allocated(const char *fname, bool set)
1412 {
1413         struct flist_head *entry;
1414         bool ret;
1415
1416         ret = file_bloom_exists(fname, set);
1417         if (!ret)
1418                 return ret;
1419
1420         flist_for_each(entry, &filename_list) {
1421                 struct file_name *fn;
1422
1423                 fn = flist_entry(entry, struct file_name, list);
1424
1425                 if (!strcmp(fn->filename, fname))
1426                         return true;
1427         }
1428
1429         return false;
1430 }
1431
1432 static bool is_already_allocated(const char *fname)
1433 {
1434         bool ret;
1435
1436         fio_file_hash_lock();
1437         ret = __is_already_allocated(fname, false);
1438         fio_file_hash_unlock();
1439
1440         return ret;
1441 }
1442
1443 static void set_already_allocated(const char *fname)
1444 {
1445         struct file_name *fn;
1446
1447         fn = malloc(sizeof(struct file_name));
1448         fn->filename = strdup(fname);
1449
1450         fio_file_hash_lock();
1451         if (!__is_already_allocated(fname, true)) {
1452                 flist_add_tail(&fn->list, &filename_list);
1453                 fn = NULL;
1454         }
1455         fio_file_hash_unlock();
1456
1457         if (fn) {
1458                 free(fn->filename);
1459                 free(fn);
1460         }
1461 }
1462
1463 static void free_already_allocated(void)
1464 {
1465         struct flist_head *entry, *tmp;
1466         struct file_name *fn;
1467
1468         if (flist_empty(&filename_list))
1469                 return;
1470
1471         fio_file_hash_lock();
1472         flist_for_each_safe(entry, tmp, &filename_list) {
1473                 fn = flist_entry(entry, struct file_name, list);
1474                 free(fn->filename);
1475                 flist_del(&fn->list);
1476                 free(fn);
1477         }
1478
1479         fio_file_hash_unlock();
1480 }
1481
1482 static struct fio_file *alloc_new_file(struct thread_data *td)
1483 {
1484         struct fio_file *f;
1485
1486         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1487                 f = calloc(1, sizeof(*f));
1488         else
1489                 f = smalloc(sizeof(*f));
1490         if (!f) {
1491                 assert(0);
1492                 return NULL;
1493         }
1494
1495         f->fd = -1;
1496         f->shadow_fd = -1;
1497         fio_file_reset(td, f);
1498         return f;
1499 }
1500
1501 bool exists_and_not_regfile(const char *filename)
1502 {
1503         struct stat sb;
1504
1505         if (lstat(filename, &sb) == -1)
1506                 return false;
1507
1508 #ifndef WIN32 /* NOT Windows */
1509         if (S_ISREG(sb.st_mode))
1510                 return false;
1511 #else
1512         /* \\.\ is the device namespace in Windows, where every file
1513          * is a device node */
1514         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1515                 return false;
1516 #endif
1517
1518         return true;
1519 }
1520
1521 static int create_work_dirs(struct thread_data *td, const char *fname)
1522 {
1523         char path[PATH_MAX];
1524         char *start, *end;
1525
1526         if (td->o.directory) {
1527                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1528                          FIO_OS_PATH_SEPARATOR, fname);
1529                 start = strstr(path, fname);
1530         } else {
1531                 snprintf(path, PATH_MAX, "%s", fname);
1532                 start = path;
1533         }
1534
1535         end = start;
1536         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1537                 if (end == start)
1538                         break;
1539                 *end = '\0';
1540                 errno = 0;
1541 #ifdef CONFIG_HAVE_MKDIR_TWO
1542                 if (mkdir(path, 0600) && errno != EEXIST) {
1543 #else
1544                 if (mkdir(path) && errno != EEXIST) {
1545 #endif
1546                         log_err("fio: failed to create dir (%s): %d\n",
1547                                 start, errno);
1548                         return 1;
1549                 }
1550                 *end = FIO_OS_PATH_SEPARATOR;
1551                 end++;
1552         }
1553         td->flags |= TD_F_DIRS_CREATED;
1554         return 0;
1555 }
1556
1557 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1558 {
1559         int cur_files = td->files_index;
1560         char file_name[PATH_MAX];
1561         struct fio_file *f;
1562         int len = 0;
1563
1564         dprint(FD_FILE, "add file %s\n", fname);
1565
1566         if (td->o.directory)
1567                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1568                                         td->o.unique_filename);
1569
1570         sprintf(file_name + len, "%s", fname);
1571
1572         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1573             !(td->flags & TD_F_DIRS_CREATED) &&
1574             create_work_dirs(td, fname))
1575                 return 1;
1576
1577         /* clean cloned siblings using existing files */
1578         if (numjob && is_already_allocated(file_name) &&
1579             !exists_and_not_regfile(fname))
1580                 return 0;
1581
1582         f = alloc_new_file(td);
1583
1584         if (td->files_size <= td->files_index) {
1585                 unsigned int new_size = td->o.nr_files + 1;
1586
1587                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1588
1589                 td->files = realloc(td->files, new_size * sizeof(f));
1590                 if (td->files == NULL) {
1591                         log_err("fio: realloc OOM\n");
1592                         assert(0);
1593                 }
1594                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1595                         td->file_locks = realloc(td->file_locks, new_size);
1596                         if (!td->file_locks) {
1597                                 log_err("fio: realloc OOM\n");
1598                                 assert(0);
1599                         }
1600                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1601                 }
1602                 td->files_size = new_size;
1603         }
1604         td->files[cur_files] = f;
1605         f->fileno = cur_files;
1606
1607         /*
1608          * init function, io engine may not be loaded yet
1609          */
1610         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1611                 f->real_file_size = -1ULL;
1612
1613         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1614                 f->file_name = strdup(file_name);
1615         else
1616                 f->file_name = smalloc_strdup(file_name);
1617         if (!f->file_name)
1618                 assert(0);
1619
1620         get_file_type(f);
1621
1622         switch (td->o.file_lock_mode) {
1623         case FILE_LOCK_NONE:
1624                 break;
1625         case FILE_LOCK_READWRITE:
1626                 f->rwlock = fio_rwlock_init();
1627                 break;
1628         case FILE_LOCK_EXCLUSIVE:
1629                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1630                 break;
1631         default:
1632                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1633                 assert(0);
1634         }
1635
1636         td->files_index++;
1637         if (f->filetype == FIO_TYPE_FILE)
1638                 td->nr_normal_files++;
1639
1640         if (td->o.numjobs > 1)
1641                 set_already_allocated(file_name);
1642
1643         if (inc)
1644                 td->o.nr_files++;
1645
1646         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1647                                                         cur_files);
1648
1649         return cur_files;
1650 }
1651
1652 int add_file_exclusive(struct thread_data *td, const char *fname)
1653 {
1654         struct fio_file *f;
1655         unsigned int i;
1656
1657         for_each_file(td, f, i) {
1658                 if (!strcmp(f->file_name, fname))
1659                         return i;
1660         }
1661
1662         return add_file(td, fname, 0, 1);
1663 }
1664
1665 void get_file(struct fio_file *f)
1666 {
1667         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1668         assert(fio_file_open(f));
1669         f->references++;
1670 }
1671
1672 int put_file(struct thread_data *td, struct fio_file *f)
1673 {
1674         int f_ret = 0, ret = 0;
1675
1676         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1677
1678         if (!fio_file_open(f)) {
1679                 assert(f->fd == -1);
1680                 return 0;
1681         }
1682
1683         assert(f->references);
1684         if (--f->references)
1685                 return 0;
1686
1687         if (should_fsync(td) && td->o.fsync_on_close) {
1688                 f_ret = fsync(f->fd);
1689                 if (f_ret < 0)
1690                         f_ret = errno;
1691         }
1692
1693         if (td->io_ops->close_file)
1694                 ret = td->io_ops->close_file(td, f);
1695
1696         if (!ret)
1697                 ret = f_ret;
1698
1699         td->nr_open_files--;
1700         fio_file_clear_open(f);
1701         assert(f->fd == -1);
1702         return ret;
1703 }
1704
1705 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1706 {
1707         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1708                 return;
1709
1710         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1711                 if (ddir == DDIR_READ)
1712                         fio_rwlock_read(f->rwlock);
1713                 else
1714                         fio_rwlock_write(f->rwlock);
1715         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1716                 fio_mutex_down(f->lock);
1717
1718         td->file_locks[f->fileno] = td->o.file_lock_mode;
1719 }
1720
1721 void unlock_file(struct thread_data *td, struct fio_file *f)
1722 {
1723         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1724                 return;
1725
1726         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1727                 fio_rwlock_unlock(f->rwlock);
1728         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1729                 fio_mutex_up(f->lock);
1730
1731         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1732 }
1733
1734 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1735 {
1736         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1737                 return;
1738         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1739                 unlock_file(td, f);
1740 }
1741
1742 static int recurse_dir(struct thread_data *td, const char *dirname)
1743 {
1744         struct dirent *dir;
1745         int ret = 0;
1746         DIR *D;
1747
1748         D = opendir(dirname);
1749         if (!D) {
1750                 char buf[FIO_VERROR_SIZE];
1751
1752                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1753                 td_verror(td, errno, buf);
1754                 return 1;
1755         }
1756
1757         while ((dir = readdir(D)) != NULL) {
1758                 char full_path[PATH_MAX];
1759                 struct stat sb;
1760
1761                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1762                         continue;
1763
1764                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1765
1766                 if (lstat(full_path, &sb) == -1) {
1767                         if (errno != ENOENT) {
1768                                 td_verror(td, errno, "stat");
1769                                 ret = 1;
1770                                 break;
1771                         }
1772                 }
1773
1774                 if (S_ISREG(sb.st_mode)) {
1775                         add_file(td, full_path, 0, 1);
1776                         continue;
1777                 }
1778                 if (!S_ISDIR(sb.st_mode))
1779                         continue;
1780
1781                 ret = recurse_dir(td, full_path);
1782                 if (ret)
1783                         break;
1784         }
1785
1786         closedir(D);
1787         return ret;
1788 }
1789
1790 int add_dir_files(struct thread_data *td, const char *path)
1791 {
1792         int ret = recurse_dir(td, path);
1793
1794         if (!ret)
1795                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1796
1797         return ret;
1798 }
1799
1800 void dup_files(struct thread_data *td, struct thread_data *org)
1801 {
1802         struct fio_file *f;
1803         unsigned int i;
1804
1805         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1806
1807         if (!org->files)
1808                 return;
1809
1810         td->files = malloc(org->files_index * sizeof(f));
1811
1812         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1813                 td->file_locks = malloc(org->files_index);
1814
1815         for_each_file(org, f, i) {
1816                 struct fio_file *__f;
1817
1818                 __f = alloc_new_file(td);
1819
1820                 if (f->file_name) {
1821                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1822                                 __f->file_name = strdup(f->file_name);
1823                         else
1824                                 __f->file_name = smalloc_strdup(f->file_name);
1825                         if (!__f->file_name)
1826                                 assert(0);
1827
1828                         __f->filetype = f->filetype;
1829                 }
1830
1831                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1832                         __f->lock = f->lock;
1833                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1834                         __f->rwlock = f->rwlock;
1835
1836                 td->files[i] = __f;
1837         }
1838 }
1839
1840 /*
1841  * Returns the index that matches the filename, or -1 if not there
1842  */
1843 int get_fileno(struct thread_data *td, const char *fname)
1844 {
1845         struct fio_file *f;
1846         unsigned int i;
1847
1848         for_each_file(td, f, i)
1849                 if (!strcmp(f->file_name, fname))
1850                         return i;
1851
1852         return -1;
1853 }
1854
1855 /*
1856  * For log usage, where we add/open/close files automatically
1857  */
1858 void free_release_files(struct thread_data *td)
1859 {
1860         close_files(td);
1861         td->o.nr_files = 0;
1862         td->o.open_files = 0;
1863         td->files_index = 0;
1864         td->nr_normal_files = 0;
1865 }
1866
1867 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1868 {
1869         int i;
1870
1871         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1872                 f->last_pos[i] = f->file_offset;
1873                 f->last_start[i] = -1ULL;
1874         }
1875
1876         if (fio_file_axmap(f))
1877                 axmap_reset(f->io_axmap);
1878         else if (fio_file_lfsr(f))
1879                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1880 }
1881
1882 bool fio_files_done(struct thread_data *td)
1883 {
1884         struct fio_file *f;
1885         unsigned int i;
1886
1887         for_each_file(td, f, i)
1888                 if (!fio_file_done(f))
1889                         return false;
1890
1891         return true;
1892 }
1893
1894 /* free memory used in initialization phase only */
1895 void filesetup_mem_free(void)
1896 {
1897         free_already_allocated();
1898 }
1899
1900 /*
1901  * This function is for platforms which support direct I/O but not O_DIRECT.
1902  */
1903 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1904 {
1905 #ifdef FIO_OS_DIRECTIO
1906         int ret = fio_set_odirect(f);
1907
1908         if (ret) {
1909                 td_verror(td, ret, "fio_set_directio");
1910 #if defined(__sun__)
1911                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1912                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1913                 } else {
1914                         log_err("fio: the file system does not seem to support direct IO\n");
1915                 }
1916 #else
1917                 log_err("fio: the file system does not seem to support direct IO\n");
1918 #endif
1919                 return -1;
1920         }
1921
1922         return 0;
1923 #else
1924         log_err("fio: direct IO is not supported on this host operating system\n");
1925         return -1;
1926 #endif
1927 }