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