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 bool pre_read_file(struct thread_data *td, struct fio_file *f)
 259 {
 260         int r, did_open = 0, old_runstate;
 261         unsigned long long left;
 262         unsigned int bs;
 263         bool ret = true;
 264         char *b;
 265
 266         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 267             td_ioengine_flagged(td, FIO_NOIO))
 268                 return true;
 269
 270         if (f->filetype == FIO_TYPE_CHAR)
 271                 return true;
 272
 273         if (!fio_file_open(f)) {
 274                 if (td->io_ops->open_file(td, f)) {
 275                         log_err("fio: cannot pre-read, failed to open file\n");
 276                         return false;
 277                 }
 278                 did_open = 1;
 279         }
 280
 281         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 282
 283         left = f->io_size;
 284         bs = td->o.max_bs[DDIR_READ];
 285         if (bs > left)
 286                 bs = left;
 287
 288         b = malloc(bs);
 289         if (!b) {
 290                 td_verror(td, errno, "malloc");
 291                 ret = false;
 292                 goto error;
 293         }
 294         memset(b, 0, bs);
 295
 296         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 297                 td_verror(td, errno, "lseek");
 298                 log_err("fio: failed to lseek pre-read file\n");
 299                 ret = false;
 300                 goto error;
 301         }
 302
 303         while (left && !td->terminate) {
 304                 if (bs > left)
 305                         bs = left;
 306
 307                 r = read(f->fd, b, bs);
 308
 309                 if (r == (int) bs) {
 310                         left -= bs;
 311                         continue;
 312                 } else {
 313                         td_verror(td, EIO, "pre_read");
 314                         break;
 315                 }
 316         }
 317
 318 error:
 319         td_restore_runstate(td, old_runstate);
 320
 321         if (did_open)
 322                 td->io_ops->close_file(td, f);
 323
 324         free(b);
 325         return ret;
 326 }
 327
 328 unsigned long long get_rand_file_size(struct thread_data *td)
 329 {
 330         unsigned long long ret, sized;
 331         uint64_t frand_max;
 332         unsigned long r;
 333
 334         frand_max = rand_max(&td->file_size_state);
 335         r = __rand(&td->file_size_state);
 336         sized = td->o.file_size_high - td->o.file_size_low;
 337         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 338         ret += td->o.file_size_low;
 339         ret -= (ret % td->o.rw_min_bs);
 340         return ret;
 341 }
 342
 343 static int file_size(struct thread_data *td, struct fio_file *f)
 344 {
 345         struct stat st;
 346
 347         if (stat(f->file_name, &st) == -1) {
 348                 td_verror(td, errno, "fstat");
 349                 return 1;
 350         }
 351
 352         f->real_file_size = st.st_size;
 353         return 0;
 354 }
 355
 356 static int bdev_size(struct thread_data *td, struct fio_file *f)
 357 {
 358         unsigned long long bytes = 0;
 359         int r;
 360
 361         if (td->io_ops->open_file(td, f)) {
 362                 log_err("fio: failed opening blockdev %s for size check\n",
 363                         f->file_name);
 364                 return 1;
 365         }
 366
 367         r = blockdev_size(f, &bytes);
 368         if (r) {
 369                 td_verror(td, r, "blockdev_size");
 370                 goto err;
 371         }
 372
 373         if (!bytes) {
 374                 log_err("%s: zero sized block device?\n", f->file_name);
 375                 goto err;
 376         }
 377
 378         f->real_file_size = bytes;
 379         td->io_ops->close_file(td, f);
 380         return 0;
 381 err:
 382         td->io_ops->close_file(td, f);
 383         return 1;
 384 }
 385
 386 static int char_size(struct thread_data *td, struct fio_file *f)
 387 {
 388 #ifdef FIO_HAVE_CHARDEV_SIZE
 389         unsigned long long bytes = 0;
 390         int r;
 391
 392         if (td->io_ops->open_file(td, f)) {
 393                 log_err("fio: failed opening chardev %s for size check\n",
 394                         f->file_name);
 395                 return 1;
 396         }
 397
 398         r = chardev_size(f, &bytes);
 399         if (r) {
 400                 td_verror(td, r, "chardev_size");
 401                 goto err;
 402         }
 403
 404         if (!bytes) {
 405                 log_err("%s: zero sized char device?\n", f->file_name);
 406                 goto err;
 407         }
 408
 409         f->real_file_size = bytes;
 410         td->io_ops->close_file(td, f);
 411         return 0;
 412 err:
 413         td->io_ops->close_file(td, f);
 414         return 1;
 415 #else
 416         f->real_file_size = -1ULL;
 417         return 0;
 418 #endif
 419 }
 420
 421 static int get_file_size(struct thread_data *td, struct fio_file *f)
 422 {
 423         int ret = 0;
 424
 425         if (fio_file_size_known(f))
 426                 return 0;
 427
 428         if (f->filetype == FIO_TYPE_FILE)
 429                 ret = file_size(td, f);
 430         else if (f->filetype == FIO_TYPE_BLOCK)
 431                 ret = bdev_size(td, f);
 432         else if (f->filetype == FIO_TYPE_CHAR)
 433                 ret = char_size(td, f);
 434         else
 435                 f->real_file_size = -1ULL;
 436
 437         /*
 438          * Leave ->real_file_size with 0 since it could be expectation
 439          * of initial setup for regular files.
 440          */
 441         if (ret)
 442                 return ret;
 443
 444         /*
 445          * If ->real_file_size is -1, a conditional for the message
 446          * "offset extends end" is always true, but it makes no sense,
 447          * so just return the same value here.
 448          */
 449         if (f->real_file_size == -1ULL) {
 450                 log_info("%s: failed to get file size of %s\n", td->o.name,
 451                                         f->file_name);
 452                 return 1;
 453         }
 454
 455         if (td->o.start_offset && f->file_offset == 0)
 456                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 457                                         f->file_name);
 458         /*
 459          * ->file_offset normally hasn't been initialized yet, so this
 460          * is basically always false.
 461          */
 462         if (f->file_offset > f->real_file_size) {
 463                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 464                                         (unsigned long long) f->file_offset,
 465                                         (unsigned long long) f->real_file_size);
 466                 return 1;
 467         }
 468
 469         fio_file_set_size_known(f);
 470         return 0;
 471 }
 472
 473 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 474                                    unsigned long long off,
 475                                    unsigned long long len)
 476 {
 477         int errval = 0, ret = 0;
 478
 479 #ifdef CONFIG_ESX
 480         return 0;
 481 #endif
 482
 483         if (len == -1ULL)
 484                 len = f->io_size;
 485         if (off == -1ULL)
 486                 off = f->file_offset;
 487
 488         if (len == -1ULL || off == -1ULL)
 489                 return 0;
 490
 491         if (td->io_ops->invalidate) {
 492                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 493                         f->file_name);
 494                 ret = td->io_ops->invalidate(td, f);
 495                 if (ret < 0)
 496                         errval = -ret;
 497         } else if (f->filetype == FIO_TYPE_FILE) {
 498                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 499                         f->file_name, off, len);
 500                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 501                 if (ret)
 502                         errval = ret;
 503         } else if (f->filetype == FIO_TYPE_BLOCK) {
 504                 int retry_count = 0;
 505
 506                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 507                 ret = blockdev_invalidate_cache(f);
 508                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 509                         /*
 510                          * Linux multipath devices reject ioctl while
 511                          * the maps are being updated. That window can
 512                          * last tens of milliseconds; we'll try up to
 513                          * a quarter of a second.
 514                          */
 515                         usleep(10000);
 516                         ret = blockdev_invalidate_cache(f);
 517                 }
 518                 if (ret < 0 && errno == EACCES && geteuid()) {
 519                         if (!root_warn) {
 520                                 log_err("fio: only root may flush block "
 521                                         "devices. Cache flush bypassed!\n");
 522                                 root_warn = 1;
 523                         }
 524                         ret = 0;
 525                 }
 526                 if (ret < 0)
 527                         errval = errno;
 528         } else if (f->filetype == FIO_TYPE_CHAR ||
 529                    f->filetype == FIO_TYPE_PIPE) {
 530                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 531                 ret = 0;
 532         }
 533
 534         /*
 535          * Cache flushing isn't a fatal condition, and we know it will
 536          * happen on some platforms where we don't have the proper
 537          * function to flush eg block device caches. So just warn and
 538          * continue on our way.
 539          */
 540         if (errval)
 541                 log_info("fio: cache invalidation of %s failed: %s\n",
 542                          f->file_name, strerror(errval));
 543
 544         return 0;
 545
 546 }
 547
 548 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 549 {
 550         if (!fio_file_open(f))
 551                 return 0;
 552
 553         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 554 }
 555
 556 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 557 {
 558         int ret = 0;
 559
 560         dprint(FD_FILE, "fd close %s\n", f->file_name);
 561
 562         remove_file_hash(f);
 563
 564         if (close(f->fd) < 0)
 565                 ret = errno;
 566
 567         f->fd = -1;
 568
 569         if (f->shadow_fd != -1) {
 570                 close(f->shadow_fd);
 571                 f->shadow_fd = -1;
 572         }
 573
 574         f->engine_pos = 0;
 575         return ret;
 576 }
 577
 578 int file_lookup_open(struct fio_file *f, int flags)
 579 {
 580         struct fio_file *__f;
 581         int from_hash;
 582
 583         __f = lookup_file_hash(f->file_name);
 584         if (__f) {
 585                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 586                 f->lock = __f->lock;
 587                 from_hash = 1;
 588         } else {
 589                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 590                 from_hash = 0;
 591         }
 592
 593 #ifdef WIN32
 594         flags |= _O_BINARY;
 595 #endif
 596
 597         f->fd = open(f->file_name, flags, 0600);
 598         return from_hash;
 599 }
 600
 601 static int file_close_shadow_fds(struct thread_data *td)
 602 {
 603         struct fio_file *f;
 604         int num_closed = 0;
 605         unsigned int i;
 606
 607         for_each_file(td, f, i) {
 608                 if (f->shadow_fd == -1)
 609                         continue;
 610
 611                 close(f->shadow_fd);
 612                 f->shadow_fd = -1;
 613                 num_closed++;
 614         }
 615
 616         return num_closed;
 617 }
 618
 619 int generic_open_file(struct thread_data *td, struct fio_file *f)
 620 {
 621         int is_std = 0;
 622         int flags = 0;
 623         int from_hash = 0;
 624
 625         dprint(FD_FILE, "fd open %s\n", f->file_name);
 626
 627         if (!strcmp(f->file_name, "-")) {
 628                 if (td_rw(td)) {
 629                         log_err("fio: can't read/write to stdin/out\n");
 630                         return 1;
 631                 }
 632                 is_std = 1;
 633
 634                 /*
 635                  * move output logging to stderr, if we are writing to stdout
 636                  */
 637                 if (td_write(td))
 638                         f_out = stderr;
 639         }
 640
 641         if (td_trim(td))
 642                 goto skip_flags;
 643         if (td->o.odirect)
 644                 flags |= OS_O_DIRECT;
 645         if (td->o.oatomic) {
 646                 if (!FIO_O_ATOMIC) {
 647                         td_verror(td, EINVAL, "OS does not support atomic IO");
 648                         return 1;
 649                 }
 650                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 651         }
 652         if (td->o.sync_io)
 653                 flags |= O_SYNC;
 654         if (td->o.create_on_open && td->o.allow_create)
 655                 flags |= O_CREAT;
 656 skip_flags:
 657         if (f->filetype != FIO_TYPE_FILE)
 658                 flags |= FIO_O_NOATIME;
 659
 660 open_again:
 661         if (td_write(td)) {
 662                 if (!read_only)
 663                         flags |= O_RDWR;
 664
 665                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 666                         flags |= O_CREAT;
 667
 668                 if (is_std)
 669                         f->fd = dup(STDOUT_FILENO);
 670                 else
 671                         from_hash = file_lookup_open(f, flags);
 672         } else if (td_read(td)) {
 673                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 674                         flags |= O_RDWR;
 675                 else
 676                         flags |= O_RDONLY;
 677
 678                 if (is_std)
 679                         f->fd = dup(STDIN_FILENO);
 680                 else
 681                         from_hash = file_lookup_open(f, flags);
 682         } else if (td_trim(td)) {
 683                 assert(!td_rw(td)); /* should have matched above */
 684                 flags |= O_RDWR;
 685                 from_hash = file_lookup_open(f, flags);
 686         }
 687
 688         if (f->fd == -1) {
 689                 char buf[FIO_VERROR_SIZE];
 690                 int __e = errno;
 691
 692                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 693                         flags &= ~FIO_O_NOATIME;
 694                         goto open_again;
 695                 }
 696                 if (__e == EMFILE && file_close_shadow_fds(td))
 697                         goto open_again;
 698
 699                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 700
 701                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 702                         log_err("fio: looks like your file system does not " \
 703                                 "support direct=1/buffered=0\n");
 704                 }
 705
 706                 td_verror(td, __e, buf);
 707                 return 1;
 708         }
 709
 710         if (!from_hash && f->fd != -1) {
 711                 if (add_file_hash(f)) {
 712                         int fio_unused ret;
 713
 714                         /*
 715                          * Stash away descriptor for later close. This is to
 716                          * work-around a "feature" on Linux, where a close of
 717                          * an fd that has been opened for write will trigger
 718                          * udev to call blkid to check partitions, fs id, etc.
 719                          * That pollutes the device cache, which can slow down
 720                          * unbuffered accesses.
 721                          */
 722                         if (f->shadow_fd == -1)
 723                                 f->shadow_fd = f->fd;
 724                         else {
 725                                 /*
 726                                  * OK to ignore, we haven't done anything
 727                                  * with it
 728                                  */
 729                                 ret = generic_close_file(td, f);
 730                         }
 731                         goto open_again;
 732                 }
 733         }
 734
 735         return 0;
 736 }
 737
 738 /*
 739  * This function i.e. get_file_size() is the default .get_file_size
 740  * implementation of majority of I/O engines.
 741  */
 742 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 743 {
 744         return get_file_size(td, f);
 745 }
 746
 747 /*
 748  * open/close all files, so that ->real_file_size gets set
 749  */
 750 static int get_file_sizes(struct thread_data *td)
 751 {
 752         struct fio_file *f;
 753         unsigned int i;
 754         int err = 0;
 755
 756         for_each_file(td, f, i) {
 757                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 758                                                                 f->file_name);
 759
 760                 if (td_io_get_file_size(td, f)) {
 761                         if (td->error != ENOENT) {
 762                                 log_err("%s\n", td->verror);
 763                                 err = 1;
 764                                 break;
 765                         }
 766                         clear_error(td);
 767                 }
 768
 769                 /*
 770                  * There are corner cases where we end up with -1 for
 771                  * ->real_file_size due to unsupported file type, etc.
 772                  * We then just set to size option value divided by number
 773                  * of files, similar to the way file ->io_size is set.
 774                  * stat(2) failure doesn't set ->real_file_size to -1.
 775                  */
 776                 if (f->real_file_size == -1ULL && td->o.size)
 777                         f->real_file_size = td->o.size / td->o.nr_files;
 778         }
 779
 780         return err;
 781 }
 782
 783 struct fio_mount {
 784         struct flist_head list;
 785         const char *base;
 786         char __base[256];
 787         unsigned int key;
 788 };
 789
 790 /*
 791  * Get free number of bytes for each file on each unique mount.
 792  */
 793 static unsigned long long get_fs_free_counts(struct thread_data *td)
 794 {
 795         struct flist_head *n, *tmp;
 796         unsigned long long ret = 0;
 797         struct fio_mount *fm;
 798         FLIST_HEAD(list);
 799         struct fio_file *f;
 800         unsigned int i;
 801
 802         for_each_file(td, f, i) {
 803                 struct stat sb;
 804                 char buf[256];
 805
 806                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 807                         if (f->real_file_size != -1ULL)
 808                                 ret += f->real_file_size;
 809                         continue;
 810                 } else if (f->filetype != FIO_TYPE_FILE)
 811                         continue;
 812
 813                 buf[255] = '\0';
 814                 strncpy(buf, f->file_name, 255);
 815
 816                 if (stat(buf, &sb) < 0) {
 817                         if (errno != ENOENT)
 818                                 break;
 819                         strcpy(buf, ".");
 820                         if (stat(buf, &sb) < 0)
 821                                 break;
 822                 }
 823
 824                 fm = NULL;
 825                 flist_for_each(n, &list) {
 826                         fm = flist_entry(n, struct fio_mount, list);
 827                         if (fm->key == sb.st_dev)
 828                                 break;
 829
 830                         fm = NULL;
 831                 }
 832
 833                 if (fm)
 834                         continue;
 835
 836                 fm = calloc(1, sizeof(*fm));
 837                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 838                 fm->base = basename(fm->__base);
 839                 fm->key = sb.st_dev;
 840                 flist_add(&fm->list, &list);
 841         }
 842
 843         flist_for_each_safe(n, tmp, &list) {
 844                 unsigned long long sz;
 845
 846                 fm = flist_entry(n, struct fio_mount, list);
 847                 flist_del(&fm->list);
 848
 849                 sz = get_fs_free_size(fm->base);
 850                 if (sz && sz != -1ULL)
 851                         ret += sz;
 852
 853                 free(fm);
 854         }
 855
 856         return ret;
 857 }
 858
 859 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 860 {
 861         struct thread_options *o = &td->o;
 862         unsigned long long align_bs;
 863         unsigned long long offset;
 864
 865         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 866                 return f->real_file_size;
 867
 868         if (o->start_offset_percent > 0) {
 869                 /*
 870                  * if offset_align is provided, set initial offset
 871                  */
 872                 if (fio_option_is_set(o, start_offset_align)) {
 873                         align_bs = o->start_offset_align;
 874                 } else {
 875                         /* else take the minimum block size */
 876                         align_bs = td_min_bs(td);
 877                 }
 878
 879                 /* calculate the raw offset */
 880                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 881                         (td->subjob_number * o->offset_increment);
 882
 883                 /*
 884                  * block align the offset at the next available boundary at
 885                  * ceiling(offset / align_bs) * align_bs
 886                  */
 887                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 888
 889         } else {
 890                 /* start_offset_percent not set */
 891                 offset = o->start_offset +
 892                                 td->subjob_number * o->offset_increment;
 893         }
 894
 895         return offset;
 896 }
 897
 898 /*
 899  * Open the files and setup files sizes, creating files if necessary.
 900  */
 901 int setup_files(struct thread_data *td)
 902 {
 903         unsigned long long total_size, extend_size;
 904         struct thread_options *o = &td->o;
 905         struct fio_file *f;
 906         unsigned int i, nr_fs_extra = 0;
 907         int err = 0, need_extend;
 908         int old_state;
 909         const unsigned int bs = td_min_bs(td);
 910         uint64_t fs = 0;
 911
 912         dprint(FD_FILE, "setup files\n");
 913
 914         old_state = td_bump_runstate(td, TD_SETTING_UP);
 915
 916         if (o->read_iolog_file)
 917                 goto done;
 918
 919         /*
 920          * Find out physical size of files or devices for this thread,
 921          * before we determine I/O size and range of our targets.
 922          * If ioengine defines a setup() method, it's responsible for
 923          * opening the files and setting f->real_file_size to indicate
 924          * the valid range for that file.
 925          */
 926         if (td->io_ops->setup)
 927                 err = td->io_ops->setup(td);
 928         else
 929                 err = get_file_sizes(td);
 930
 931         if (err)
 932                 goto err_out;
 933
 934         /*
 935          * check sizes. if the files/devices do not exist and the size
 936          * isn't passed to fio, abort.
 937          */
 938         total_size = 0;
 939         for_each_file(td, f, i) {
 940                 f->fileno = i;
 941                 if (f->real_file_size == -1ULL)
 942                         total_size = -1ULL;
 943                 else
 944                         total_size += f->real_file_size;
 945         }
 946
 947         if (o->fill_device)
 948                 td->fill_device_size = get_fs_free_counts(td);
 949
 950         /*
 951          * device/file sizes are zero and no size given, punt
 952          */
 953         if ((!total_size || total_size == -1ULL) && !o->size &&
 954             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 955             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 956                 log_err("%s: you need to specify size=\n", o->name);
 957                 td_verror(td, EINVAL, "total_file_size");
 958                 goto err_out;
 959         }
 960
 961         /*
 962          * Calculate per-file size and potential extra size for the
 963          * first files, if needed (i.e. if we don't have a fixed size).
 964          */
 965         if (!o->file_size_low && o->nr_files) {
 966                 uint64_t all_fs;
 967
 968                 fs = o->size / o->nr_files;
 969                 all_fs = fs * o->nr_files;
 970
 971                 if (all_fs < o->size)
 972                         nr_fs_extra = (o->size - all_fs) / bs;
 973         }
 974
 975         /*
 976          * now file sizes are known, so we can set ->io_size. if size= is
 977          * not given, ->io_size is just equal to ->real_file_size. if size
 978          * is given, ->io_size is size / nr_files.
 979          */
 980         extend_size = total_size = 0;
 981         need_extend = 0;
 982         for_each_file(td, f, i) {
 983                 f->file_offset = get_start_offset(td, f);
 984
 985                 /*
 986                  * Update ->io_size depending on options specified.
 987                  * ->file_size_low being 0 means filesize option isn't set.
 988                  * Non zero ->file_size_low equals ->file_size_high means
 989                  * filesize option is set in a fixed size format.
 990                  * Non zero ->file_size_low not equals ->file_size_high means
 991                  * filesize option is set in a range format.
 992                  */
 993                 if (!o->file_size_low) {
 994                         /*
 995                          * no file size or range given, file size is equal to
 996                          * total size divided by number of files. If the size
 997                          * doesn't divide nicely with the min blocksize,
 998                          * make the first files bigger.
 999                          */
1000                         f->io_size = fs;
1001                         if (nr_fs_extra) {
1002                                 nr_fs_extra--;
1003                                 f->io_size += bs;
1004                         }
1005
1006                         /*
1007                          * We normally don't come here for regular files, but
1008                          * if the result is 0 for a regular file, set it to the
1009                          * real file size. This could be size of the existing
1010                          * one if it already exists, but otherwise will be set
1011                          * to 0. A new file won't be created because
1012                          * ->io_size + ->file_offset equals ->real_file_size.
1013                          */
1014                         if (!f->io_size) {
1015                                 if (f->file_offset > f->real_file_size)
1016                                         goto err_offset;
1017                                 f->io_size = f->real_file_size - f->file_offset;
1018                                 if (!f->io_size)
1019                                         log_info("fio: file %s may be ignored\n",
1020                                                 f->file_name);
1021                         }
1022                 } else if (f->real_file_size < o->file_size_low ||
1023                            f->real_file_size > o->file_size_high) {
1024                         if (f->file_offset > o->file_size_low)
1025                                 goto err_offset;
1026                         /*
1027                          * file size given. if it's fixed, use that. if it's a
1028                          * range, generate a random size in-between.
1029                          */
1030                         if (o->file_size_low == o->file_size_high)
1031                                 f->io_size = o->file_size_low - f->file_offset;
1032                         else {
1033                                 f->io_size = get_rand_file_size(td)
1034                                                 - f->file_offset;
1035                         }
1036                 } else
1037                         f->io_size = f->real_file_size - f->file_offset;
1038
1039                 if (f->io_size == -1ULL)
1040                         total_size = -1ULL;
1041                 else {
1042                         if (o->size_percent) {
1043                                 uint64_t file_size;
1044
1045                                 file_size = f->io_size + f->file_offset;
1046                                 f->io_size = (file_size *
1047                                               o->size_percent) / 100;
1048                                 if (f->io_size > (file_size - f->file_offset))
1049                                         f->io_size = file_size - f->file_offset;
1050
1051                                 f->io_size -= (f->io_size % td_min_bs(td));
1052                         }
1053                         total_size += f->io_size;
1054                 }
1055
1056                 if (f->filetype == FIO_TYPE_FILE &&
1057                     (f->io_size + f->file_offset) > f->real_file_size &&
1058                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1059                         if (!o->create_on_open) {
1060                                 need_extend++;
1061                                 extend_size += (f->io_size + f->file_offset);
1062                                 fio_file_set_extend(f);
1063                         } else
1064                                 f->real_file_size = f->io_size + f->file_offset;
1065                 }
1066         }
1067
1068         if (td->o.block_error_hist) {
1069                 int len;
1070
1071                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1072                 f = td->files[0];
1073                 len = f->io_size / td->o.bs[DDIR_TRIM];
1074                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1075                         log_err("fio: cannot calculate block histogram with "
1076                                 "%d trim blocks, maximum %d\n",
1077                                 len, MAX_NR_BLOCK_INFOS);
1078                         td_verror(td, EINVAL, "block_error_hist");
1079                         goto err_out;
1080                 }
1081
1082                 td->ts.nr_block_infos = len;
1083                 for (i = 0; i < len; i++)
1084                         td->ts.block_infos[i] =
1085                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1086         } else
1087                 td->ts.nr_block_infos = 0;
1088
1089         if (!o->size || (total_size && o->size > total_size))
1090                 o->size = total_size;
1091
1092         if (o->size < td_min_bs(td)) {
1093                 log_err("fio: blocksize too large for data set\n");
1094                 goto err_out;
1095         }
1096
1097         /*
1098          * See if we need to extend some files, typically needed when our
1099          * target regular files don't exist yet, but our jobs require them
1100          * initially due to read I/Os.
1101          */
1102         if (need_extend) {
1103                 temp_stall_ts = 1;
1104                 if (output_format & FIO_OUTPUT_NORMAL) {
1105                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1106                                  o->name,
1107                                  need_extend > 1 ? "s" : "",
1108                                  need_extend,
1109                                  need_extend > 1 ? "s" : "",
1110                                  need_extend > 1 ? "total " : "",
1111                                  extend_size >> 20);
1112                 }
1113
1114                 for_each_file(td, f, i) {
1115                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1116
1117                         if (!fio_file_extend(f))
1118                                 continue;
1119
1120                         assert(f->filetype == FIO_TYPE_FILE);
1121                         fio_file_clear_extend(f);
1122                         if (!o->fill_device) {
1123                                 old_len = f->real_file_size;
1124                                 extend_len = f->io_size + f->file_offset -
1125                                                 old_len;
1126                         }
1127                         f->real_file_size = (f->io_size + f->file_offset);
1128                         err = extend_file(td, f);
1129                         if (err)
1130                                 break;
1131
1132                         err = __file_invalidate_cache(td, f, old_len,
1133                                                                 extend_len);
1134
1135                         /*
1136                          * Shut up static checker
1137                          */
1138                         if (f->fd != -1)
1139                                 close(f->fd);
1140
1141                         f->fd = -1;
1142                         if (err)
1143                                 break;
1144                 }
1145                 temp_stall_ts = 0;
1146         }
1147
1148         if (err)
1149                 goto err_out;
1150
1151         if (!o->zone_size)
1152                 o->zone_size = o->size;
1153
1154         /*
1155          * iolog already set the total io size, if we read back
1156          * stored entries.
1157          */
1158         if (!o->read_iolog_file) {
1159                 if (o->io_size)
1160                         td->total_io_size = o->io_size * o->loops;
1161                 else
1162                         td->total_io_size = o->size * o->loops;
1163         }
1164
1165 done:
1166         if (o->create_only)
1167                 td->done = 1;
1168
1169         td_restore_runstate(td, old_state);
1170         return 0;
1171 err_offset:
1172         log_err("%s: you need to specify valid offset=\n", o->name);
1173 err_out:
1174         td_restore_runstate(td, old_state);
1175         return 1;
1176 }
1177
1178 bool pre_read_files(struct thread_data *td)
1179 {
1180         struct fio_file *f;
1181         unsigned int i;
1182
1183         dprint(FD_FILE, "pre_read files\n");
1184
1185         for_each_file(td, f, i) {
1186                 if (!pre_read_file(td, f))
1187                         return false;
1188         }
1189
1190         return true;
1191 }
1192
1193 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1194 {
1195         unsigned int range_size, seed;
1196         unsigned long nranges;
1197         uint64_t fsize;
1198
1199         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1200         fsize = min(f->real_file_size, f->io_size);
1201
1202         nranges = (fsize + range_size - 1) / range_size;
1203
1204         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1205         if (!td->o.rand_repeatable)
1206                 seed = td->rand_seeds[4];
1207
1208         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1209                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1210         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1211                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1212         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1213                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1214
1215         return 1;
1216 }
1217
1218 static int init_rand_distribution(struct thread_data *td)
1219 {
1220         struct fio_file *f;
1221         unsigned int i;
1222         int state;
1223
1224         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1225                 return 0;
1226
1227         state = td_bump_runstate(td, TD_SETTING_UP);
1228
1229         for_each_file(td, f, i)
1230                 __init_rand_distribution(td, f);
1231
1232         td_restore_runstate(td, state);
1233
1234         return 1;
1235 }
1236
1237 /*
1238  * Check if the number of blocks exceeds the randomness capability of
1239  * the selected generator. Tausworthe is 32-bit, the others are fullly
1240  * 64-bit capable.
1241  */
1242 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1243                                  uint64_t blocks)
1244 {
1245         if (blocks <= FRAND32_MAX)
1246                 return 0;
1247         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1248                 return 0;
1249
1250         /*
1251          * If the user hasn't specified a random generator, switch
1252          * to tausworthe64 with informational warning. If the user did
1253          * specify one, just warn.
1254          */
1255         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1256                         f->file_name);
1257
1258         if (!fio_option_is_set(&td->o, random_generator)) {
1259                 log_info("fio: Switching to tausworthe64. Use the "
1260                          "random_generator= option to get rid of this "
1261                          "warning.\n");
1262                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1263                 return 0;
1264         }
1265
1266         /*
1267          * Just make this information to avoid breaking scripts.
1268          */
1269         log_info("fio: Use the random_generator= option to switch to lfsr or "
1270                          "tausworthe64.\n");
1271         return 0;
1272 }
1273
1274 int init_random_map(struct thread_data *td)
1275 {
1276         unsigned long long blocks;
1277         struct fio_file *f;
1278         unsigned int i;
1279
1280         if (init_rand_distribution(td))
1281                 return 0;
1282         if (!td_random(td))
1283                 return 0;
1284
1285         for_each_file(td, f, i) {
1286                 uint64_t fsize = min(f->real_file_size, f->io_size);
1287
1288                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1289
1290                 if (check_rand_gen_limits(td, f, blocks))
1291                         return 1;
1292
1293                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1294                         unsigned long seed;
1295
1296                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1297
1298                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1299                                 fio_file_set_lfsr(f);
1300                                 continue;
1301                         }
1302                 } else if (!td->o.norandommap) {
1303                         f->io_axmap = axmap_new(blocks);
1304                         if (f->io_axmap) {
1305                                 fio_file_set_axmap(f);
1306                                 continue;
1307                         }
1308                 } else if (td->o.norandommap)
1309                         continue;
1310
1311                 if (!td->o.softrandommap) {
1312                         log_err("fio: failed allocating random map. If running"
1313                                 " a large number of jobs, try the 'norandommap'"
1314                                 " option or set 'softrandommap'. Or give"
1315                                 " a larger --alloc-size to fio.\n");
1316                         return 1;
1317                 }
1318
1319                 log_info("fio: file %s failed allocating random map. Running "
1320                          "job without.\n", f->file_name);
1321         }
1322
1323         return 0;
1324 }
1325
1326 void close_files(struct thread_data *td)
1327 {
1328         struct fio_file *f;
1329         unsigned int i;
1330
1331         for_each_file(td, f, i) {
1332                 if (fio_file_open(f))
1333                         td_io_close_file(td, f);
1334         }
1335 }
1336
1337 void close_and_free_files(struct thread_data *td)
1338 {
1339         struct fio_file *f;
1340         unsigned int i;
1341         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1342
1343         dprint(FD_FILE, "close files\n");
1344
1345         for_each_file(td, f, i) {
1346                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1347                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1348                         td_io_unlink_file(td, f);
1349                 }
1350
1351                 if (fio_file_open(f))
1352                         td_io_close_file(td, f);
1353
1354                 remove_file_hash(f);
1355
1356                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1357                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1358                         td_io_unlink_file(td, f);
1359                 }
1360
1361                 if (use_free)
1362                         free(f->file_name);
1363                 else
1364                         sfree(f->file_name);
1365                 f->file_name = NULL;
1366                 if (fio_file_axmap(f)) {
1367                         axmap_free(f->io_axmap);
1368                         f->io_axmap = NULL;
1369                 }
1370                 if (use_free)
1371                         free(f);
1372                 else
1373                         sfree(f);
1374         }
1375
1376         td->o.filename = NULL;
1377         free(td->files);
1378         free(td->file_locks);
1379         td->files_index = 0;
1380         td->files = NULL;
1381         td->file_locks = NULL;
1382         td->o.file_lock_mode = FILE_LOCK_NONE;
1383         td->o.nr_files = 0;
1384 }
1385
1386 static void get_file_type(struct fio_file *f)
1387 {
1388         struct stat sb;
1389
1390         if (!strcmp(f->file_name, "-"))
1391                 f->filetype = FIO_TYPE_PIPE;
1392         else
1393                 f->filetype = FIO_TYPE_FILE;
1394
1395 #ifdef WIN32
1396         /* \\.\ is the device namespace in Windows, where every file is
1397          * a block device */
1398         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1399                 f->filetype = FIO_TYPE_BLOCK;
1400 #endif
1401
1402         if (!stat(f->file_name, &sb)) {
1403                 if (S_ISBLK(sb.st_mode))
1404                         f->filetype = FIO_TYPE_BLOCK;
1405                 else if (S_ISCHR(sb.st_mode))
1406                         f->filetype = FIO_TYPE_CHAR;
1407                 else if (S_ISFIFO(sb.st_mode))
1408                         f->filetype = FIO_TYPE_PIPE;
1409         }
1410 }
1411
1412 static bool __is_already_allocated(const char *fname, bool set)
1413 {
1414         struct flist_head *entry;
1415         bool ret;
1416
1417         ret = file_bloom_exists(fname, set);
1418         if (!ret)
1419                 return ret;
1420
1421         flist_for_each(entry, &filename_list) {
1422                 struct file_name *fn;
1423
1424                 fn = flist_entry(entry, struct file_name, list);
1425
1426                 if (!strcmp(fn->filename, fname))
1427                         return true;
1428         }
1429
1430         return false;
1431 }
1432
1433 static bool is_already_allocated(const char *fname)
1434 {
1435         bool ret;
1436
1437         fio_file_hash_lock();
1438         ret = __is_already_allocated(fname, false);
1439         fio_file_hash_unlock();
1440
1441         return ret;
1442 }
1443
1444 static void set_already_allocated(const char *fname)
1445 {
1446         struct file_name *fn;
1447
1448         fn = malloc(sizeof(struct file_name));
1449         fn->filename = strdup(fname);
1450
1451         fio_file_hash_lock();
1452         if (!__is_already_allocated(fname, true)) {
1453                 flist_add_tail(&fn->list, &filename_list);
1454                 fn = NULL;
1455         }
1456         fio_file_hash_unlock();
1457
1458         if (fn) {
1459                 free(fn->filename);
1460                 free(fn);
1461         }
1462 }
1463
1464 static void free_already_allocated(void)
1465 {
1466         struct flist_head *entry, *tmp;
1467         struct file_name *fn;
1468
1469         if (flist_empty(&filename_list))
1470                 return;
1471
1472         fio_file_hash_lock();
1473         flist_for_each_safe(entry, tmp, &filename_list) {
1474                 fn = flist_entry(entry, struct file_name, list);
1475                 free(fn->filename);
1476                 flist_del(&fn->list);
1477                 free(fn);
1478         }
1479
1480         fio_file_hash_unlock();
1481 }
1482
1483 static struct fio_file *alloc_new_file(struct thread_data *td)
1484 {
1485         struct fio_file *f;
1486
1487         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1488                 f = calloc(1, sizeof(*f));
1489         else
1490                 f = smalloc(sizeof(*f));
1491         if (!f) {
1492                 assert(0);
1493                 return NULL;
1494         }
1495
1496         f->fd = -1;
1497         f->shadow_fd = -1;
1498         fio_file_reset(td, f);
1499         return f;
1500 }
1501
1502 bool exists_and_not_regfile(const char *filename)
1503 {
1504         struct stat sb;
1505
1506         if (lstat(filename, &sb) == -1)
1507                 return false;
1508
1509 #ifndef WIN32 /* NOT Windows */
1510         if (S_ISREG(sb.st_mode))
1511                 return false;
1512 #else
1513         /* \\.\ is the device namespace in Windows, where every file
1514          * is a device node */
1515         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1516                 return false;
1517 #endif
1518
1519         return true;
1520 }
1521
1522 static int create_work_dirs(struct thread_data *td, const char *fname)
1523 {
1524         char path[PATH_MAX];
1525         char *start, *end;
1526
1527         if (td->o.directory) {
1528                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1529                          FIO_OS_PATH_SEPARATOR, fname);
1530                 start = strstr(path, fname);
1531         } else {
1532                 snprintf(path, PATH_MAX, "%s", fname);
1533                 start = path;
1534         }
1535
1536         end = start;
1537         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1538                 if (end == start)
1539                         break;
1540                 *end = '\0';
1541                 errno = 0;
1542 #ifdef CONFIG_HAVE_MKDIR_TWO
1543                 if (mkdir(path, 0600) && errno != EEXIST) {
1544 #else
1545                 if (mkdir(path) && errno != EEXIST) {
1546 #endif
1547                         log_err("fio: failed to create dir (%s): %d\n",
1548                                 start, errno);
1549                         return 1;
1550                 }
1551                 *end = FIO_OS_PATH_SEPARATOR;
1552                 end++;
1553         }
1554         td->flags |= TD_F_DIRS_CREATED;
1555         return 0;
1556 }
1557
1558 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1559 {
1560         int cur_files = td->files_index;
1561         char file_name[PATH_MAX];
1562         struct fio_file *f;
1563         int len = 0;
1564
1565         dprint(FD_FILE, "add file %s\n", fname);
1566
1567         if (td->o.directory)
1568                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1569                                         td->o.unique_filename);
1570
1571         sprintf(file_name + len, "%s", fname);
1572
1573         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1574             !(td->flags & TD_F_DIRS_CREATED) &&
1575             create_work_dirs(td, fname))
1576                 return 1;
1577
1578         /* clean cloned siblings using existing files */
1579         if (numjob && is_already_allocated(file_name) &&
1580             !exists_and_not_regfile(fname))
1581                 return 0;
1582
1583         f = alloc_new_file(td);
1584
1585         if (td->files_size <= td->files_index) {
1586                 unsigned int new_size = td->o.nr_files + 1;
1587
1588                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1589
1590                 td->files = realloc(td->files, new_size * sizeof(f));
1591                 if (td->files == NULL) {
1592                         log_err("fio: realloc OOM\n");
1593                         assert(0);
1594                 }
1595                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1596                         td->file_locks = realloc(td->file_locks, new_size);
1597                         if (!td->file_locks) {
1598                                 log_err("fio: realloc OOM\n");
1599                                 assert(0);
1600                         }
1601                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1602                 }
1603                 td->files_size = new_size;
1604         }
1605         td->files[cur_files] = f;
1606         f->fileno = cur_files;
1607
1608         /*
1609          * init function, io engine may not be loaded yet
1610          */
1611         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1612                 f->real_file_size = -1ULL;
1613
1614         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1615                 f->file_name = strdup(file_name);
1616         else
1617                 f->file_name = smalloc_strdup(file_name);
1618         if (!f->file_name)
1619                 assert(0);
1620
1621         get_file_type(f);
1622
1623         switch (td->o.file_lock_mode) {
1624         case FILE_LOCK_NONE:
1625                 break;
1626         case FILE_LOCK_READWRITE:
1627                 f->rwlock = fio_rwlock_init();
1628                 break;
1629         case FILE_LOCK_EXCLUSIVE:
1630                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1631                 break;
1632         default:
1633                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1634                 assert(0);
1635         }
1636
1637         td->files_index++;
1638         if (f->filetype == FIO_TYPE_FILE)
1639                 td->nr_normal_files++;
1640
1641         if (td->o.numjobs > 1)
1642                 set_already_allocated(file_name);
1643
1644         if (inc)
1645                 td->o.nr_files++;
1646
1647         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1648                                                         cur_files);
1649
1650         return cur_files;
1651 }
1652
1653 int add_file_exclusive(struct thread_data *td, const char *fname)
1654 {
1655         struct fio_file *f;
1656         unsigned int i;
1657
1658         for_each_file(td, f, i) {
1659                 if (!strcmp(f->file_name, fname))
1660                         return i;
1661         }
1662
1663         return add_file(td, fname, 0, 1);
1664 }
1665
1666 void get_file(struct fio_file *f)
1667 {
1668         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1669         assert(fio_file_open(f));
1670         f->references++;
1671 }
1672
1673 int put_file(struct thread_data *td, struct fio_file *f)
1674 {
1675         int f_ret = 0, ret = 0;
1676
1677         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1678
1679         if (!fio_file_open(f)) {
1680                 assert(f->fd == -1);
1681                 return 0;
1682         }
1683
1684         assert(f->references);
1685         if (--f->references)
1686                 return 0;
1687
1688         if (should_fsync(td) && td->o.fsync_on_close) {
1689                 f_ret = fsync(f->fd);
1690                 if (f_ret < 0)
1691                         f_ret = errno;
1692         }
1693
1694         if (td->io_ops->close_file)
1695                 ret = td->io_ops->close_file(td, f);
1696
1697         if (!ret)
1698                 ret = f_ret;
1699
1700         td->nr_open_files--;
1701         fio_file_clear_open(f);
1702         assert(f->fd == -1);
1703         return ret;
1704 }
1705
1706 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1707 {
1708         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1709                 return;
1710
1711         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1712                 if (ddir == DDIR_READ)
1713                         fio_rwlock_read(f->rwlock);
1714                 else
1715                         fio_rwlock_write(f->rwlock);
1716         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1717                 fio_mutex_down(f->lock);
1718
1719         td->file_locks[f->fileno] = td->o.file_lock_mode;
1720 }
1721
1722 void unlock_file(struct thread_data *td, struct fio_file *f)
1723 {
1724         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1725                 return;
1726
1727         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1728                 fio_rwlock_unlock(f->rwlock);
1729         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1730                 fio_mutex_up(f->lock);
1731
1732         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1733 }
1734
1735 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1736 {
1737         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1738                 return;
1739         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1740                 unlock_file(td, f);
1741 }
1742
1743 static int recurse_dir(struct thread_data *td, const char *dirname)
1744 {
1745         struct dirent *dir;
1746         int ret = 0;
1747         DIR *D;
1748
1749         D = opendir(dirname);
1750         if (!D) {
1751                 char buf[FIO_VERROR_SIZE];
1752
1753                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1754                 td_verror(td, errno, buf);
1755                 return 1;
1756         }
1757
1758         while ((dir = readdir(D)) != NULL) {
1759                 char full_path[PATH_MAX];
1760                 struct stat sb;
1761
1762                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1763                         continue;
1764
1765                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1766
1767                 if (lstat(full_path, &sb) == -1) {
1768                         if (errno != ENOENT) {
1769                                 td_verror(td, errno, "stat");
1770                                 ret = 1;
1771                                 break;
1772                         }
1773                 }
1774
1775                 if (S_ISREG(sb.st_mode)) {
1776                         add_file(td, full_path, 0, 1);
1777                         continue;
1778                 }
1779                 if (!S_ISDIR(sb.st_mode))
1780                         continue;
1781
1782                 ret = recurse_dir(td, full_path);
1783                 if (ret)
1784                         break;
1785         }
1786
1787         closedir(D);
1788         return ret;
1789 }
1790
1791 int add_dir_files(struct thread_data *td, const char *path)
1792 {
1793         int ret = recurse_dir(td, path);
1794
1795         if (!ret)
1796                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1797
1798         return ret;
1799 }
1800
1801 void dup_files(struct thread_data *td, struct thread_data *org)
1802 {
1803         struct fio_file *f;
1804         unsigned int i;
1805
1806         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1807
1808         if (!org->files)
1809                 return;
1810
1811         td->files = malloc(org->files_index * sizeof(f));
1812
1813         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1814                 td->file_locks = malloc(org->files_index);
1815
1816         for_each_file(org, f, i) {
1817                 struct fio_file *__f;
1818
1819                 __f = alloc_new_file(td);
1820
1821                 if (f->file_name) {
1822                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1823                                 __f->file_name = strdup(f->file_name);
1824                         else
1825                                 __f->file_name = smalloc_strdup(f->file_name);
1826                         if (!__f->file_name)
1827                                 assert(0);
1828
1829                         __f->filetype = f->filetype;
1830                 }
1831
1832                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1833                         __f->lock = f->lock;
1834                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1835                         __f->rwlock = f->rwlock;
1836
1837                 td->files[i] = __f;
1838         }
1839 }
1840
1841 /*
1842  * Returns the index that matches the filename, or -1 if not there
1843  */
1844 int get_fileno(struct thread_data *td, const char *fname)
1845 {
1846         struct fio_file *f;
1847         unsigned int i;
1848
1849         for_each_file(td, f, i)
1850                 if (!strcmp(f->file_name, fname))
1851                         return i;
1852
1853         return -1;
1854 }
1855
1856 /*
1857  * For log usage, where we add/open/close files automatically
1858  */
1859 void free_release_files(struct thread_data *td)
1860 {
1861         close_files(td);
1862         td->o.nr_files = 0;
1863         td->o.open_files = 0;
1864         td->files_index = 0;
1865         td->nr_normal_files = 0;
1866 }
1867
1868 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1869 {
1870         int i;
1871
1872         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1873                 f->last_pos[i] = f->file_offset;
1874                 f->last_start[i] = -1ULL;
1875         }
1876
1877         if (fio_file_axmap(f))
1878                 axmap_reset(f->io_axmap);
1879         else if (fio_file_lfsr(f))
1880                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1881 }
1882
1883 bool fio_files_done(struct thread_data *td)
1884 {
1885         struct fio_file *f;
1886         unsigned int i;
1887
1888         for_each_file(td, f, i)
1889                 if (!fio_file_done(f))
1890                         return false;
1891
1892         return true;
1893 }
1894
1895 /* free memory used in initialization phase only */
1896 void filesetup_mem_free(void)
1897 {
1898         free_already_allocated();
1899 }
1900
1901 /*
1902  * This function is for platforms which support direct I/O but not O_DIRECT.
1903  */
1904 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1905 {
1906 #ifdef FIO_OS_DIRECTIO
1907         int ret = fio_set_odirect(f);
1908
1909         if (ret) {
1910                 td_verror(td, ret, "fio_set_directio");
1911 #if defined(__sun__)
1912                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1913                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1914                 } else {
1915                         log_err("fio: the file system does not seem to support direct IO\n");
1916                 }
1917 #else
1918                 log_err("fio: the file system does not seem to support direct IO\n");
1919 #endif
1920                 return -1;
1921         }
1922
1923         return 0;
1924 #else
1925         log_err("fio: direct IO is not supported on this host operating system\n");
1926         return -1;
1927 #endif
1928 }