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 #include "lib/memalign.h"
  19
  20 #ifdef CONFIG_LINUX_FALLOCATE
  21 #include <linux/falloc.h>
  22 #endif
  23
  24 static int root_warn;
  25
  26 static FLIST_HEAD(filename_list);
  27
  28 /*
  29  * List entry for filename_list
  30  */
  31 struct file_name {
  32         struct flist_head list;
  33         char *filename;
  34 };
  35
  36 static inline void clear_error(struct thread_data *td)
  37 {
  38         td->error = 0;
  39         td->verror[0] = '\0';
  40 }
  41
  42 static inline int native_fallocate(struct thread_data *td, struct fio_file *f)
  43 {
  44         bool success;
  45
  46         success = fio_fallocate(f, 0, f->real_file_size);
  47         dprint(FD_FILE, "native fallocate of file %s size %llu was "
  48                         "%ssuccessful\n", f->file_name,
  49                         (unsigned long long) f->real_file_size,
  50                         !success ? "un": "");
  51
  52         if (success)
  53                 return 0;
  54
  55         if (errno == ENOSYS)
  56                 dprint(FD_FILE, "native fallocate is not implemented\n");
  57
  58         return -1;
  59 }
  60
  61 static void fallocate_file(struct thread_data *td, struct fio_file *f)
  62 {
  63         int r;
  64
  65         if (td->o.fill_device)
  66                 return;
  67
  68         switch (td->o.fallocate_mode) {
  69         case FIO_FALLOCATE_NATIVE:
  70                 r = native_fallocate(td, f);
  71                 if (r != 0 && errno != ENOSYS)
  72                         log_err("fio: native_fallocate call failed: %s\n",
  73                                         strerror(errno));
  74                 break;
  75         case FIO_FALLOCATE_NONE:
  76                 break;
  77 #ifdef CONFIG_POSIX_FALLOCATE
  78         case FIO_FALLOCATE_POSIX:
  79                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  80                                  f->file_name,
  81                                  (unsigned long long) f->real_file_size);
  82
  83                 r = posix_fallocate(f->fd, 0, f->real_file_size);
  84                 if (r > 0)
  85                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
  86                 break;
  87 #endif /* CONFIG_POSIX_FALLOCATE */
  88 #ifdef CONFIG_LINUX_FALLOCATE
  89         case FIO_FALLOCATE_KEEP_SIZE:
  90                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
  91                                 "file %s size %llu\n", f->file_name,
  92                                 (unsigned long long) f->real_file_size);
  93
  94                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
  95                 if (r != 0)
  96                         td_verror(td, errno, "fallocate");
  97
  98                 break;
  99 #endif /* CONFIG_LINUX_FALLOCATE */
 100         default:
 101                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
 102                 assert(0);
 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         if (td->o.odirect)
 151                 flags |= OS_O_DIRECT;
 152
 153 #ifdef WIN32
 154         flags |= _O_BINARY;
 155 #endif
 156
 157         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 158         f->fd = open(f->file_name, flags, 0644);
 159         if (f->fd < 0) {
 160                 int err = errno;
 161
 162                 if (err == ENOENT && !td->o.allow_create)
 163                         log_err("fio: file creation disallowed by "
 164                                         "allow_file_create=0\n");
 165                 else {
 166                         if (err == EINVAL && (flags & OS_O_DIRECT))
 167                                 log_err("fio: looks like your filesystem "
 168                                         "does not support "
 169                                         "direct=1/buffered=0\n");
 170
 171                         td_verror(td, err, "open");
 172                 }
 173                 return 1;
 174         }
 175
 176         fallocate_file(td, f);
 177
 178         /*
 179          * If our jobs don't require regular files initially, we're done.
 180          */
 181         if (!new_layout)
 182                 goto done;
 183
 184         /*
 185          * The size will be -1ULL when fill_device is used, so don't truncate
 186          * or fallocate this file, just write it
 187          */
 188         if (!td->o.fill_device) {
 189                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 190                                         (unsigned long long) f->real_file_size);
 191                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 192                         if (errno != EFBIG) {
 193                                 td_verror(td, errno, "ftruncate");
 194                                 goto err;
 195                         }
 196                 }
 197         }
 198
 199         if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
 200                 goto err;
 201
 202         left = f->real_file_size;
 203         bs = td->o.max_bs[DDIR_WRITE];
 204         if (bs > left)
 205                 bs = left;
 206
 207         b = fio_memalign(page_size, bs);
 208         if (!b) {
 209                 td_verror(td, errno, "fio_memalign");
 210                 goto err;
 211         }
 212
 213         while (left && !td->terminate) {
 214                 ssize_t r;
 215
 216                 if (bs > left)
 217                         bs = left;
 218
 219                 fill_io_buffer(td, b, bs, bs);
 220
 221                 r = write(f->fd, b, bs);
 222
 223                 if (r > 0) {
 224                         left -= r;
 225                         continue;
 226                 } else {
 227                         if (r < 0) {
 228                                 int __e = errno;
 229
 230                                 if (__e == ENOSPC) {
 231                                         if (td->o.fill_device)
 232                                                 break;
 233                                         log_info("fio: ENOSPC on laying out "
 234                                                  "file, stopping\n");
 235                                         break;
 236                                 }
 237                                 td_verror(td, errno, "write");
 238                         } else
 239                                 td_verror(td, EIO, "write");
 240
 241                         break;
 242                 }
 243         }
 244
 245         if (td->terminate) {
 246                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 247                 td_io_unlink_file(td, f);
 248         } else if (td->o.create_fsync) {
 249                 if (fsync(f->fd) < 0) {
 250                         td_verror(td, errno, "fsync");
 251                         goto err;
 252                 }
 253         }
 254         if (td->o.fill_device && !td_write(td)) {
 255                 fio_file_clear_size_known(f);
 256                 if (td_io_get_file_size(td, f))
 257                         goto err;
 258                 if (f->io_size > f->real_file_size)
 259                         f->io_size = f->real_file_size;
 260         }
 261
 262         fio_memfree(b, bs);
 263 done:
 264         return 0;
 265 err:
 266         close(f->fd);
 267         f->fd = -1;
 268         if (b)
 269                 fio_memfree(b, bs);
 270         return 1;
 271 }
 272
 273 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 274 {
 275         int ret = 0, r, did_open = 0, old_runstate;
 276         unsigned long long left;
 277         unsigned int bs;
 278         char *b;
 279
 280         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 281             td_ioengine_flagged(td, FIO_NOIO))
 282                 return 0;
 283
 284         if (f->filetype == FIO_TYPE_CHAR)
 285                 return 0;
 286
 287         if (!fio_file_open(f)) {
 288                 if (td->io_ops->open_file(td, f)) {
 289                         log_err("fio: cannot pre-read, failed to open file\n");
 290                         return 1;
 291                 }
 292                 did_open = 1;
 293         }
 294
 295         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 296
 297         left = f->io_size;
 298         bs = td->o.max_bs[DDIR_READ];
 299         if (bs > left)
 300                 bs = left;
 301
 302         b = malloc(bs);
 303         if (!b) {
 304                 td_verror(td, errno, "malloc");
 305                 ret = 1;
 306                 goto error;
 307         }
 308         memset(b, 0, bs);
 309
 310         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 311                 td_verror(td, errno, "lseek");
 312                 log_err("fio: failed to lseek pre-read file\n");
 313                 ret = 1;
 314                 goto error;
 315         }
 316
 317         while (left && !td->terminate) {
 318                 if (bs > left)
 319                         bs = left;
 320
 321                 r = read(f->fd, b, bs);
 322
 323                 if (r == (int) bs) {
 324                         left -= bs;
 325                         continue;
 326                 } else {
 327                         td_verror(td, EIO, "pre_read");
 328                         break;
 329                 }
 330         }
 331
 332 error:
 333         td_restore_runstate(td, old_runstate);
 334
 335         if (did_open)
 336                 td->io_ops->close_file(td, f);
 337
 338         free(b);
 339         return ret;
 340 }
 341
 342 unsigned long long get_rand_file_size(struct thread_data *td)
 343 {
 344         unsigned long long ret, sized;
 345         uint64_t frand_max;
 346         unsigned long r;
 347
 348         frand_max = rand_max(&td->file_size_state);
 349         r = __rand(&td->file_size_state);
 350         sized = td->o.file_size_high - td->o.file_size_low;
 351         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 352         ret += td->o.file_size_low;
 353         ret -= (ret % td->o.rw_min_bs);
 354         return ret;
 355 }
 356
 357 static int file_size(struct thread_data *td, struct fio_file *f)
 358 {
 359         struct stat st;
 360
 361         if (stat(f->file_name, &st) == -1) {
 362                 td_verror(td, errno, "fstat");
 363                 return 1;
 364         }
 365
 366         f->real_file_size = st.st_size;
 367         return 0;
 368 }
 369
 370 static int bdev_size(struct thread_data *td, struct fio_file *f)
 371 {
 372         unsigned long long bytes = 0;
 373         int r;
 374
 375         if (td->io_ops->open_file(td, f)) {
 376                 log_err("fio: failed opening blockdev %s for size check\n",
 377                         f->file_name);
 378                 return 1;
 379         }
 380
 381         r = blockdev_size(f, &bytes);
 382         if (r) {
 383                 td_verror(td, r, "blockdev_size");
 384                 goto err;
 385         }
 386
 387         if (!bytes) {
 388                 log_err("%s: zero sized block device?\n", f->file_name);
 389                 goto err;
 390         }
 391
 392         f->real_file_size = bytes;
 393         td->io_ops->close_file(td, f);
 394         return 0;
 395 err:
 396         td->io_ops->close_file(td, f);
 397         return 1;
 398 }
 399
 400 static int char_size(struct thread_data *td, struct fio_file *f)
 401 {
 402 #ifdef FIO_HAVE_CHARDEV_SIZE
 403         unsigned long long bytes = 0;
 404         int r;
 405
 406         if (td->io_ops->open_file(td, f)) {
 407                 log_err("fio: failed opening chardev %s for size check\n",
 408                         f->file_name);
 409                 return 1;
 410         }
 411
 412         r = chardev_size(f, &bytes);
 413         if (r) {
 414                 td_verror(td, r, "chardev_size");
 415                 goto err;
 416         }
 417
 418         if (!bytes) {
 419                 log_err("%s: zero sized char device?\n", f->file_name);
 420                 goto err;
 421         }
 422
 423         f->real_file_size = bytes;
 424         td->io_ops->close_file(td, f);
 425         return 0;
 426 err:
 427         td->io_ops->close_file(td, f);
 428         return 1;
 429 #else
 430         f->real_file_size = -1ULL;
 431         return 0;
 432 #endif
 433 }
 434
 435 static int get_file_size(struct thread_data *td, struct fio_file *f)
 436 {
 437         int ret = 0;
 438
 439         if (fio_file_size_known(f))
 440                 return 0;
 441
 442         if (f->filetype == FIO_TYPE_FILE)
 443                 ret = file_size(td, f);
 444         else if (f->filetype == FIO_TYPE_BLOCK)
 445                 ret = bdev_size(td, f);
 446         else if (f->filetype == FIO_TYPE_CHAR)
 447                 ret = char_size(td, f);
 448         else
 449                 f->real_file_size = -1ULL;
 450
 451         /*
 452          * Leave ->real_file_size with 0 since it could be expectation
 453          * of initial setup for regular files.
 454          */
 455         if (ret)
 456                 return ret;
 457
 458         /*
 459          * If ->real_file_size is -1, a conditional for the message
 460          * "offset extends end" is always true, but it makes no sense,
 461          * so just return the same value here.
 462          */
 463         if (f->real_file_size == -1ULL) {
 464                 log_info("%s: failed to get file size of %s\n", td->o.name,
 465                                         f->file_name);
 466                 return 1;
 467         }
 468
 469         if (td->o.start_offset && f->file_offset == 0)
 470                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 471                                         f->file_name);
 472         /*
 473          * ->file_offset normally hasn't been initialized yet, so this
 474          * is basically always false.
 475          */
 476         if (f->file_offset > f->real_file_size) {
 477                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 478                                         (unsigned long long) f->file_offset,
 479                                         (unsigned long long) f->real_file_size);
 480                 return 1;
 481         }
 482
 483         fio_file_set_size_known(f);
 484         return 0;
 485 }
 486
 487 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 488                                    unsigned long long off,
 489                                    unsigned long long len)
 490 {
 491         int errval = 0, ret = 0;
 492
 493 #ifdef CONFIG_ESX
 494         return 0;
 495 #endif
 496
 497         if (len == -1ULL)
 498                 len = f->io_size;
 499         if (off == -1ULL)
 500                 off = f->file_offset;
 501
 502         if (len == -1ULL || off == -1ULL)
 503                 return 0;
 504
 505         if (td->io_ops->invalidate) {
 506                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 507                         f->file_name);
 508                 ret = td->io_ops->invalidate(td, f);
 509                 if (ret < 0)
 510                         errval = -ret;
 511         } else if (f->filetype == FIO_TYPE_FILE) {
 512                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 513                         f->file_name, off, len);
 514                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 515                 if (ret)
 516                         errval = ret;
 517         } else if (f->filetype == FIO_TYPE_BLOCK) {
 518                 int retry_count = 0;
 519
 520                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 521                 ret = blockdev_invalidate_cache(f);
 522                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 523                         /*
 524                          * Linux multipath devices reject ioctl while
 525                          * the maps are being updated. That window can
 526                          * last tens of milliseconds; we'll try up to
 527                          * a quarter of a second.
 528                          */
 529                         usleep(10000);
 530                         ret = blockdev_invalidate_cache(f);
 531                 }
 532                 if (ret < 0 && errno == EACCES && geteuid()) {
 533                         if (!root_warn) {
 534                                 log_err("fio: only root may flush block "
 535                                         "devices. Cache flush bypassed!\n");
 536                                 root_warn = 1;
 537                         }
 538                         ret = 0;
 539                 }
 540                 if (ret < 0)
 541                         errval = errno;
 542         } else if (f->filetype == FIO_TYPE_CHAR ||
 543                    f->filetype == FIO_TYPE_PIPE) {
 544                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 545                 ret = 0;
 546         }
 547
 548         /*
 549          * Cache flushing isn't a fatal condition, and we know it will
 550          * happen on some platforms where we don't have the proper
 551          * function to flush eg block device caches. So just warn and
 552          * continue on our way.
 553          */
 554         if (errval)
 555                 log_info("fio: cache invalidation of %s failed: %s\n",
 556                          f->file_name, strerror(errval));
 557
 558         return 0;
 559
 560 }
 561
 562 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 563 {
 564         if (!fio_file_open(f))
 565                 return 0;
 566
 567         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 568 }
 569
 570 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 571 {
 572         int ret = 0;
 573
 574         dprint(FD_FILE, "fd close %s\n", f->file_name);
 575
 576         remove_file_hash(f);
 577
 578         if (close(f->fd) < 0)
 579                 ret = errno;
 580
 581         f->fd = -1;
 582
 583         if (f->shadow_fd != -1) {
 584                 close(f->shadow_fd);
 585                 f->shadow_fd = -1;
 586         }
 587
 588         f->engine_pos = 0;
 589         return ret;
 590 }
 591
 592 int file_lookup_open(struct fio_file *f, int flags)
 593 {
 594         struct fio_file *__f;
 595         int from_hash;
 596
 597         __f = lookup_file_hash(f->file_name);
 598         if (__f) {
 599                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 600                 f->lock = __f->lock;
 601                 from_hash = 1;
 602         } else {
 603                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 604                 from_hash = 0;
 605         }
 606
 607 #ifdef WIN32
 608         flags |= _O_BINARY;
 609 #endif
 610
 611         f->fd = open(f->file_name, flags, 0600);
 612         return from_hash;
 613 }
 614
 615 static int file_close_shadow_fds(struct thread_data *td)
 616 {
 617         struct fio_file *f;
 618         int num_closed = 0;
 619         unsigned int i;
 620
 621         for_each_file(td, f, i) {
 622                 if (f->shadow_fd == -1)
 623                         continue;
 624
 625                 close(f->shadow_fd);
 626                 f->shadow_fd = -1;
 627                 num_closed++;
 628         }
 629
 630         return num_closed;
 631 }
 632
 633 int generic_open_file(struct thread_data *td, struct fio_file *f)
 634 {
 635         int is_std = 0;
 636         int flags = 0;
 637         int from_hash = 0;
 638
 639         dprint(FD_FILE, "fd open %s\n", f->file_name);
 640
 641         if (!strcmp(f->file_name, "-")) {
 642                 if (td_rw(td)) {
 643                         log_err("fio: can't read/write to stdin/out\n");
 644                         return 1;
 645                 }
 646                 is_std = 1;
 647
 648                 /*
 649                  * move output logging to stderr, if we are writing to stdout
 650                  */
 651                 if (td_write(td))
 652                         f_out = stderr;
 653         }
 654
 655         if (td_trim(td))
 656                 goto skip_flags;
 657         if (td->o.odirect)
 658                 flags |= OS_O_DIRECT;
 659         if (td->o.oatomic) {
 660                 if (!FIO_O_ATOMIC) {
 661                         td_verror(td, EINVAL, "OS does not support atomic IO");
 662                         return 1;
 663                 }
 664                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 665         }
 666         if (td->o.sync_io)
 667                 flags |= O_SYNC;
 668         if (td->o.create_on_open && td->o.allow_create)
 669                 flags |= O_CREAT;
 670 skip_flags:
 671         if (f->filetype != FIO_TYPE_FILE)
 672                 flags |= FIO_O_NOATIME;
 673
 674 open_again:
 675         if (td_write(td)) {
 676                 if (!read_only)
 677                         flags |= O_RDWR;
 678
 679                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 680                         flags |= O_CREAT;
 681
 682                 if (is_std)
 683                         f->fd = dup(STDOUT_FILENO);
 684                 else
 685                         from_hash = file_lookup_open(f, flags);
 686         } else if (td_read(td)) {
 687                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 688                         flags |= O_RDWR;
 689                 else
 690                         flags |= O_RDONLY;
 691
 692                 if (is_std)
 693                         f->fd = dup(STDIN_FILENO);
 694                 else
 695                         from_hash = file_lookup_open(f, flags);
 696         } else if (td_trim(td)) {
 697                 assert(!td_rw(td)); /* should have matched above */
 698                 flags |= O_RDWR;
 699                 from_hash = file_lookup_open(f, flags);
 700         }
 701
 702         if (f->fd == -1) {
 703                 char buf[FIO_VERROR_SIZE];
 704                 int __e = errno;
 705
 706                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 707                         flags &= ~FIO_O_NOATIME;
 708                         goto open_again;
 709                 }
 710                 if (__e == EMFILE && file_close_shadow_fds(td))
 711                         goto open_again;
 712
 713                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 714
 715                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 716                         log_err("fio: looks like your file system does not " \
 717                                 "support direct=1/buffered=0\n");
 718                 }
 719
 720                 td_verror(td, __e, buf);
 721                 return 1;
 722         }
 723
 724         if (!from_hash && f->fd != -1) {
 725                 if (add_file_hash(f)) {
 726                         int fio_unused ret;
 727
 728                         /*
 729                          * Stash away descriptor for later close. This is to
 730                          * work-around a "feature" on Linux, where a close of
 731                          * an fd that has been opened for write will trigger
 732                          * udev to call blkid to check partitions, fs id, etc.
 733                          * That pollutes the device cache, which can slow down
 734                          * unbuffered accesses.
 735                          */
 736                         if (f->shadow_fd == -1)
 737                                 f->shadow_fd = f->fd;
 738                         else {
 739                                 /*
 740                                  * OK to ignore, we haven't done anything
 741                                  * with it
 742                                  */
 743                                 ret = generic_close_file(td, f);
 744                         }
 745                         goto open_again;
 746                 }
 747         }
 748
 749         return 0;
 750 }
 751
 752 /*
 753  * This function i.e. get_file_size() is the default .get_file_size
 754  * implementation of majority of I/O engines.
 755  */
 756 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 757 {
 758         return get_file_size(td, f);
 759 }
 760
 761 /*
 762  * open/close all files, so that ->real_file_size gets set
 763  */
 764 static int get_file_sizes(struct thread_data *td)
 765 {
 766         struct fio_file *f;
 767         unsigned int i;
 768         int err = 0;
 769
 770         for_each_file(td, f, i) {
 771                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 772                                                                 f->file_name);
 773
 774                 if (td_io_get_file_size(td, f)) {
 775                         if (td->error != ENOENT) {
 776                                 log_err("%s\n", td->verror);
 777                                 err = 1;
 778                                 break;
 779                         }
 780                         clear_error(td);
 781                 }
 782
 783                 /*
 784                  * There are corner cases where we end up with -1 for
 785                  * ->real_file_size due to unsupported file type, etc.
 786                  * We then just set to size option value divided by number
 787                  * of files, similar to the way file ->io_size is set.
 788                  * stat(2) failure doesn't set ->real_file_size to -1.
 789                  */
 790                 if (f->real_file_size == -1ULL && td->o.size)
 791                         f->real_file_size = td->o.size / td->o.nr_files;
 792         }
 793
 794         return err;
 795 }
 796
 797 struct fio_mount {
 798         struct flist_head list;
 799         const char *base;
 800         char __base[256];
 801         unsigned int key;
 802 };
 803
 804 /*
 805  * Get free number of bytes for each file on each unique mount.
 806  */
 807 static unsigned long long get_fs_free_counts(struct thread_data *td)
 808 {
 809         struct flist_head *n, *tmp;
 810         unsigned long long ret = 0;
 811         struct fio_mount *fm;
 812         FLIST_HEAD(list);
 813         struct fio_file *f;
 814         unsigned int i;
 815
 816         for_each_file(td, f, i) {
 817                 struct stat sb;
 818                 char buf[256];
 819
 820                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 821                         if (f->real_file_size != -1ULL)
 822                                 ret += f->real_file_size;
 823                         continue;
 824                 } else if (f->filetype != FIO_TYPE_FILE)
 825                         continue;
 826
 827                 buf[255] = '\0';
 828                 strncpy(buf, f->file_name, 255);
 829
 830                 if (stat(buf, &sb) < 0) {
 831                         if (errno != ENOENT)
 832                                 break;
 833                         strcpy(buf, ".");
 834                         if (stat(buf, &sb) < 0)
 835                                 break;
 836                 }
 837
 838                 fm = NULL;
 839                 flist_for_each(n, &list) {
 840                         fm = flist_entry(n, struct fio_mount, list);
 841                         if (fm->key == sb.st_dev)
 842                                 break;
 843
 844                         fm = NULL;
 845                 }
 846
 847                 if (fm)
 848                         continue;
 849
 850                 fm = calloc(1, sizeof(*fm));
 851                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 852                 fm->base = basename(fm->__base);
 853                 fm->key = sb.st_dev;
 854                 flist_add(&fm->list, &list);
 855         }
 856
 857         flist_for_each_safe(n, tmp, &list) {
 858                 unsigned long long sz;
 859
 860                 fm = flist_entry(n, struct fio_mount, list);
 861                 flist_del(&fm->list);
 862
 863                 sz = get_fs_free_size(fm->base);
 864                 if (sz && sz != -1ULL)
 865                         ret += sz;
 866
 867                 free(fm);
 868         }
 869
 870         return ret;
 871 }
 872
 873 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 874 {
 875         struct thread_options *o = &td->o;
 876         unsigned long long align_bs;
 877         unsigned long long offset;
 878
 879         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 880                 return f->real_file_size;
 881
 882         if (o->start_offset_percent > 0) {
 883                 /*
 884                  * if blockalign is provided, find the min across read, write,
 885                  * and trim
 886                  */
 887                 if (fio_option_is_set(o, ba)) {
 888                         align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
 889                         align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
 890                 } else {
 891                         /* else take the minimum block size */
 892                         align_bs = td_min_bs(td);
 893                 }
 894
 895                 /* calculate the raw offset */
 896                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 897                         (td->subjob_number * o->offset_increment);
 898
 899                 /*
 900                  * block align the offset at the next available boundary at
 901                  * ceiling(offset / align_bs) * align_bs
 902                  */
 903                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 904
 905         } else {
 906                 /* start_offset_percent not set */
 907                 offset = o->start_offset +
 908                                 td->subjob_number * o->offset_increment;
 909         }
 910
 911         return offset;
 912 }
 913
 914 /*
 915  * Open the files and setup files sizes, creating files if necessary.
 916  */
 917 int setup_files(struct thread_data *td)
 918 {
 919         unsigned long long total_size, extend_size;
 920         struct thread_options *o = &td->o;
 921         struct fio_file *f;
 922         unsigned int i, nr_fs_extra = 0;
 923         int err = 0, need_extend;
 924         int old_state;
 925         const unsigned int bs = td_min_bs(td);
 926         uint64_t fs = 0;
 927
 928         dprint(FD_FILE, "setup files\n");
 929
 930         old_state = td_bump_runstate(td, TD_SETTING_UP);
 931
 932         if (o->read_iolog_file)
 933                 goto done;
 934
 935         /*
 936          * Find out physical size of files or devices for this thread,
 937          * before we determine I/O size and range of our targets.
 938          * If ioengine defines a setup() method, it's responsible for
 939          * opening the files and setting f->real_file_size to indicate
 940          * the valid range for that file.
 941          */
 942         if (td->io_ops->setup)
 943                 err = td->io_ops->setup(td);
 944         else
 945                 err = get_file_sizes(td);
 946
 947         if (err)
 948                 goto err_out;
 949
 950         /*
 951          * check sizes. if the files/devices do not exist and the size
 952          * isn't passed to fio, abort.
 953          */
 954         total_size = 0;
 955         for_each_file(td, f, i) {
 956                 f->fileno = i;
 957                 if (f->real_file_size == -1ULL)
 958                         total_size = -1ULL;
 959                 else
 960                         total_size += f->real_file_size;
 961         }
 962
 963         if (o->fill_device)
 964                 td->fill_device_size = get_fs_free_counts(td);
 965
 966         /*
 967          * device/file sizes are zero and no size given, punt
 968          */
 969         if ((!total_size || total_size == -1ULL) && !o->size &&
 970             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 971             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 972                 log_err("%s: you need to specify size=\n", o->name);
 973                 td_verror(td, EINVAL, "total_file_size");
 974                 goto err_out;
 975         }
 976
 977         /*
 978          * Calculate per-file size and potential extra size for the
 979          * first files, if needed (i.e. if we don't have a fixed size).
 980          */
 981         if (!o->file_size_low && o->nr_files) {
 982                 uint64_t all_fs;
 983
 984                 fs = o->size / o->nr_files;
 985                 all_fs = fs * o->nr_files;
 986
 987                 if (all_fs < o->size)
 988                         nr_fs_extra = (o->size - all_fs) / bs;
 989         }
 990
 991         /*
 992          * now file sizes are known, so we can set ->io_size. if size= is
 993          * not given, ->io_size is just equal to ->real_file_size. if size
 994          * is given, ->io_size is size / nr_files.
 995          */
 996         extend_size = total_size = 0;
 997         need_extend = 0;
 998         for_each_file(td, f, i) {
 999                 f->file_offset = get_start_offset(td, f);
1000
1001                 /*
1002                  * Update ->io_size depending on options specified.
1003                  * ->file_size_low being 0 means filesize option isn't set.
1004                  * Non zero ->file_size_low equals ->file_size_high means
1005                  * filesize option is set in a fixed size format.
1006                  * Non zero ->file_size_low not equals ->file_size_high means
1007                  * filesize option is set in a range format.
1008                  */
1009                 if (!o->file_size_low) {
1010                         /*
1011                          * no file size or range given, file size is equal to
1012                          * total size divided by number of files. If the size
1013                          * doesn't divide nicely with the min blocksize,
1014                          * make the first files bigger.
1015                          */
1016                         f->io_size = fs;
1017                         if (nr_fs_extra) {
1018                                 nr_fs_extra--;
1019                                 f->io_size += bs;
1020                         }
1021
1022                         /*
1023                          * We normally don't come here for regular files, but
1024                          * if the result is 0 for a regular file, set it to the
1025                          * real file size. This could be size of the existing
1026                          * one if it already exists, but otherwise will be set
1027                          * to 0. A new file won't be created because
1028                          * ->io_size + ->file_offset equals ->real_file_size.
1029                          */
1030                         if (!f->io_size) {
1031                                 if (f->file_offset > f->real_file_size)
1032                                         goto err_offset;
1033                                 f->io_size = f->real_file_size - f->file_offset;
1034                                 if (!f->io_size)
1035                                         log_info("fio: file %s may be ignored\n",
1036                                                 f->file_name);
1037                         }
1038                 } else if (f->real_file_size < o->file_size_low ||
1039                            f->real_file_size > o->file_size_high) {
1040                         if (f->file_offset > o->file_size_low)
1041                                 goto err_offset;
1042                         /*
1043                          * file size given. if it's fixed, use that. if it's a
1044                          * range, generate a random size in-between.
1045                          */
1046                         if (o->file_size_low == o->file_size_high)
1047                                 f->io_size = o->file_size_low - f->file_offset;
1048                         else {
1049                                 f->io_size = get_rand_file_size(td)
1050                                                 - f->file_offset;
1051                         }
1052                 } else
1053                         f->io_size = f->real_file_size - f->file_offset;
1054
1055                 if (f->io_size == -1ULL)
1056                         total_size = -1ULL;
1057                 else {
1058                         if (o->size_percent) {
1059                                 uint64_t file_size;
1060
1061                                 file_size = f->io_size + f->file_offset;
1062                                 f->io_size = (file_size *
1063                                               o->size_percent) / 100;
1064                                 if (f->io_size > (file_size - f->file_offset))
1065                                         f->io_size = file_size - f->file_offset;
1066
1067                                 f->io_size -= (f->io_size % td_min_bs(td));
1068                         }
1069                         total_size += f->io_size;
1070                 }
1071
1072                 if (f->filetype == FIO_TYPE_FILE &&
1073                     (f->io_size + f->file_offset) > f->real_file_size &&
1074                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1075                         if (!o->create_on_open) {
1076                                 need_extend++;
1077                                 extend_size += (f->io_size + f->file_offset);
1078                                 fio_file_set_extend(f);
1079                         } else
1080                                 f->real_file_size = f->io_size + f->file_offset;
1081                 }
1082         }
1083
1084         if (td->o.block_error_hist) {
1085                 int len;
1086
1087                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1088                 f = td->files[0];
1089                 len = f->io_size / td->o.bs[DDIR_TRIM];
1090                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1091                         log_err("fio: cannot calculate block histogram with "
1092                                 "%d trim blocks, maximum %d\n",
1093                                 len, MAX_NR_BLOCK_INFOS);
1094                         td_verror(td, EINVAL, "block_error_hist");
1095                         goto err_out;
1096                 }
1097
1098                 td->ts.nr_block_infos = len;
1099                 for (i = 0; i < len; i++)
1100                         td->ts.block_infos[i] =
1101                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1102         } else
1103                 td->ts.nr_block_infos = 0;
1104
1105         if (!o->size || (total_size && o->size > total_size))
1106                 o->size = total_size;
1107
1108         if (o->size < td_min_bs(td)) {
1109                 log_err("fio: blocksize too large for data set\n");
1110                 goto err_out;
1111         }
1112
1113         /*
1114          * See if we need to extend some files, typically needed when our
1115          * target regular files don't exist yet, but our jobs require them
1116          * initially due to read I/Os.
1117          */
1118         if (need_extend) {
1119                 temp_stall_ts = 1;
1120                 if (output_format & FIO_OUTPUT_NORMAL) {
1121                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1122                                  o->name,
1123                                  need_extend > 1 ? "s" : "",
1124                                  need_extend,
1125                                  need_extend > 1 ? "s" : "",
1126                                  need_extend > 1 ? "total " : "",
1127                                  extend_size >> 20);
1128                 }
1129
1130                 for_each_file(td, f, i) {
1131                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1132
1133                         if (!fio_file_extend(f))
1134                                 continue;
1135
1136                         assert(f->filetype == FIO_TYPE_FILE);
1137                         fio_file_clear_extend(f);
1138                         if (!o->fill_device) {
1139                                 old_len = f->real_file_size;
1140                                 extend_len = f->io_size + f->file_offset -
1141                                                 old_len;
1142                         }
1143                         f->real_file_size = (f->io_size + f->file_offset);
1144                         err = extend_file(td, f);
1145                         if (err)
1146                                 break;
1147
1148                         err = __file_invalidate_cache(td, f, old_len,
1149                                                                 extend_len);
1150
1151                         /*
1152                          * Shut up static checker
1153                          */
1154                         if (f->fd != -1)
1155                                 close(f->fd);
1156
1157                         f->fd = -1;
1158                         if (err)
1159                                 break;
1160                 }
1161                 temp_stall_ts = 0;
1162         }
1163
1164         if (err)
1165                 goto err_out;
1166
1167         if (!o->zone_size)
1168                 o->zone_size = o->size;
1169
1170         /*
1171          * iolog already set the total io size, if we read back
1172          * stored entries.
1173          */
1174         if (!o->read_iolog_file) {
1175                 if (o->io_size)
1176                         td->total_io_size = o->io_size * o->loops;
1177                 else
1178                         td->total_io_size = o->size * o->loops;
1179         }
1180
1181 done:
1182         if (o->create_only)
1183                 td->done = 1;
1184
1185         td_restore_runstate(td, old_state);
1186         return 0;
1187 err_offset:
1188         log_err("%s: you need to specify valid offset=\n", o->name);
1189 err_out:
1190         td_restore_runstate(td, old_state);
1191         return 1;
1192 }
1193
1194 int pre_read_files(struct thread_data *td)
1195 {
1196         struct fio_file *f;
1197         unsigned int i;
1198
1199         dprint(FD_FILE, "pre_read files\n");
1200
1201         for_each_file(td, f, i) {
1202                 if (pre_read_file(td, f))
1203                         return -1;
1204         }
1205
1206         return 0;
1207 }
1208
1209 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1210 {
1211         unsigned int range_size, seed;
1212         unsigned long nranges;
1213         uint64_t fsize;
1214
1215         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1216         fsize = min(f->real_file_size, f->io_size);
1217
1218         nranges = (fsize + range_size - 1) / range_size;
1219
1220         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1221         if (!td->o.rand_repeatable)
1222                 seed = td->rand_seeds[4];
1223
1224         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1225                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1226         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1227                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1228         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1229                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1230
1231         return 1;
1232 }
1233
1234 static int init_rand_distribution(struct thread_data *td)
1235 {
1236         struct fio_file *f;
1237         unsigned int i;
1238         int state;
1239
1240         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1241                 return 0;
1242
1243         state = td_bump_runstate(td, TD_SETTING_UP);
1244
1245         for_each_file(td, f, i)
1246                 __init_rand_distribution(td, f);
1247
1248         td_restore_runstate(td, state);
1249
1250         return 1;
1251 }
1252
1253 /*
1254  * Check if the number of blocks exceeds the randomness capability of
1255  * the selected generator. Tausworthe is 32-bit, the others are fullly
1256  * 64-bit capable.
1257  */
1258 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1259                                  uint64_t blocks)
1260 {
1261         if (blocks <= FRAND32_MAX)
1262                 return 0;
1263         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1264                 return 0;
1265
1266         /*
1267          * If the user hasn't specified a random generator, switch
1268          * to tausworthe64 with informational warning. If the user did
1269          * specify one, just warn.
1270          */
1271         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1272                         f->file_name);
1273
1274         if (!fio_option_is_set(&td->o, random_generator)) {
1275                 log_info("fio: Switching to tausworthe64. Use the "
1276                          "random_generator= option to get rid of this "
1277                          "warning.\n");
1278                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1279                 return 0;
1280         }
1281
1282         /*
1283          * Just make this information to avoid breaking scripts.
1284          */
1285         log_info("fio: Use the random_generator= option to switch to lfsr or "
1286                          "tausworthe64.\n");
1287         return 0;
1288 }
1289
1290 int init_random_map(struct thread_data *td)
1291 {
1292         unsigned long long blocks;
1293         struct fio_file *f;
1294         unsigned int i;
1295
1296         if (init_rand_distribution(td))
1297                 return 0;
1298         if (!td_random(td))
1299                 return 0;
1300
1301         for_each_file(td, f, i) {
1302                 uint64_t fsize = min(f->real_file_size, f->io_size);
1303
1304                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1305
1306                 if (check_rand_gen_limits(td, f, blocks))
1307                         return 1;
1308
1309                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1310                         unsigned long seed;
1311
1312                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1313
1314                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1315                                 fio_file_set_lfsr(f);
1316                                 continue;
1317                         }
1318                 } else if (!td->o.norandommap) {
1319                         f->io_axmap = axmap_new(blocks);
1320                         if (f->io_axmap) {
1321                                 fio_file_set_axmap(f);
1322                                 continue;
1323                         }
1324                 } else if (td->o.norandommap)
1325                         continue;
1326
1327                 if (!td->o.softrandommap) {
1328                         log_err("fio: failed allocating random map. If running"
1329                                 " a large number of jobs, try the 'norandommap'"
1330                                 " option or set 'softrandommap'. Or give"
1331                                 " a larger --alloc-size to fio.\n");
1332                         return 1;
1333                 }
1334
1335                 log_info("fio: file %s failed allocating random map. Running "
1336                          "job without.\n", f->file_name);
1337         }
1338
1339         return 0;
1340 }
1341
1342 void close_files(struct thread_data *td)
1343 {
1344         struct fio_file *f;
1345         unsigned int i;
1346
1347         for_each_file(td, f, i) {
1348                 if (fio_file_open(f))
1349                         td_io_close_file(td, f);
1350         }
1351 }
1352
1353 void close_and_free_files(struct thread_data *td)
1354 {
1355         struct fio_file *f;
1356         unsigned int i;
1357
1358         dprint(FD_FILE, "close files\n");
1359
1360         for_each_file(td, f, i) {
1361                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1362                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1363                         td_io_unlink_file(td, f);
1364                 }
1365
1366                 if (fio_file_open(f))
1367                         td_io_close_file(td, f);
1368
1369                 remove_file_hash(f);
1370
1371                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1372                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1373                         td_io_unlink_file(td, f);
1374                 }
1375
1376                 sfree(f->file_name);
1377                 f->file_name = NULL;
1378                 if (fio_file_axmap(f)) {
1379                         axmap_free(f->io_axmap);
1380                         f->io_axmap = NULL;
1381                 }
1382                 sfree(f);
1383         }
1384
1385         td->o.filename = NULL;
1386         free(td->files);
1387         free(td->file_locks);
1388         td->files_index = 0;
1389         td->files = NULL;
1390         td->file_locks = NULL;
1391         td->o.file_lock_mode = FILE_LOCK_NONE;
1392         td->o.nr_files = 0;
1393 }
1394
1395 static void get_file_type(struct fio_file *f)
1396 {
1397         struct stat sb;
1398
1399         if (!strcmp(f->file_name, "-"))
1400                 f->filetype = FIO_TYPE_PIPE;
1401         else
1402                 f->filetype = FIO_TYPE_FILE;
1403
1404 #ifdef WIN32
1405         /* \\.\ is the device namespace in Windows, where every file is
1406          * a block device */
1407         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1408                 f->filetype = FIO_TYPE_BLOCK;
1409 #endif
1410
1411         if (!stat(f->file_name, &sb)) {
1412                 if (S_ISBLK(sb.st_mode))
1413                         f->filetype = FIO_TYPE_BLOCK;
1414                 else if (S_ISCHR(sb.st_mode))
1415                         f->filetype = FIO_TYPE_CHAR;
1416                 else if (S_ISFIFO(sb.st_mode))
1417                         f->filetype = FIO_TYPE_PIPE;
1418         }
1419 }
1420
1421 static bool __is_already_allocated(const char *fname, bool set)
1422 {
1423         struct flist_head *entry;
1424         bool ret;
1425
1426         ret = file_bloom_exists(fname, set);
1427         if (!ret)
1428                 return ret;
1429
1430         flist_for_each(entry, &filename_list) {
1431                 struct file_name *fn;
1432
1433                 fn = flist_entry(entry, struct file_name, list);
1434
1435                 if (!strcmp(fn->filename, fname))
1436                         return true;
1437         }
1438
1439         return false;
1440 }
1441
1442 static bool is_already_allocated(const char *fname)
1443 {
1444         bool ret;
1445
1446         fio_file_hash_lock();
1447         ret = __is_already_allocated(fname, false);
1448         fio_file_hash_unlock();
1449
1450         return ret;
1451 }
1452
1453 static void set_already_allocated(const char *fname)
1454 {
1455         struct file_name *fn;
1456
1457         fn = malloc(sizeof(struct file_name));
1458         fn->filename = strdup(fname);
1459
1460         fio_file_hash_lock();
1461         if (!__is_already_allocated(fname, true)) {
1462                 flist_add_tail(&fn->list, &filename_list);
1463                 fn = NULL;
1464         }
1465         fio_file_hash_unlock();
1466
1467         if (fn) {
1468                 free(fn->filename);
1469                 free(fn);
1470         }
1471 }
1472
1473 static void free_already_allocated(void)
1474 {
1475         struct flist_head *entry, *tmp;
1476         struct file_name *fn;
1477
1478         if (flist_empty(&filename_list))
1479                 return;
1480
1481         fio_file_hash_lock();
1482         flist_for_each_safe(entry, tmp, &filename_list) {
1483                 fn = flist_entry(entry, struct file_name, list);
1484                 free(fn->filename);
1485                 flist_del(&fn->list);
1486                 free(fn);
1487         }
1488
1489         fio_file_hash_unlock();
1490 }
1491
1492 static struct fio_file *alloc_new_file(struct thread_data *td)
1493 {
1494         struct fio_file *f;
1495
1496         f = smalloc(sizeof(*f));
1497         if (!f) {
1498                 assert(0);
1499                 return NULL;
1500         }
1501
1502         f->fd = -1;
1503         f->shadow_fd = -1;
1504         fio_file_reset(td, f);
1505         return f;
1506 }
1507
1508 bool exists_and_not_regfile(const char *filename)
1509 {
1510         struct stat sb;
1511
1512         if (lstat(filename, &sb) == -1)
1513                 return false;
1514
1515 #ifndef WIN32 /* NOT Windows */
1516         if (S_ISREG(sb.st_mode))
1517                 return false;
1518 #else
1519         /* \\.\ is the device namespace in Windows, where every file
1520          * is a device node */
1521         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1522                 return false;
1523 #endif
1524
1525         return true;
1526 }
1527
1528 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1529 {
1530         int cur_files = td->files_index;
1531         char file_name[PATH_MAX];
1532         struct fio_file *f;
1533         int len = 0;
1534
1535         dprint(FD_FILE, "add file %s\n", fname);
1536
1537         if (td->o.directory)
1538                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1539                                         td->o.unique_filename);
1540
1541         sprintf(file_name + len, "%s", fname);
1542
1543         /* clean cloned siblings using existing files */
1544         if (numjob && is_already_allocated(file_name) &&
1545             !exists_and_not_regfile(fname))
1546                 return 0;
1547
1548         f = alloc_new_file(td);
1549
1550         if (td->files_size <= td->files_index) {
1551                 unsigned int new_size = td->o.nr_files + 1;
1552
1553                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1554
1555                 td->files = realloc(td->files, new_size * sizeof(f));
1556                 if (td->files == NULL) {
1557                         log_err("fio: realloc OOM\n");
1558                         assert(0);
1559                 }
1560                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1561                         td->file_locks = realloc(td->file_locks, new_size);
1562                         if (!td->file_locks) {
1563                                 log_err("fio: realloc OOM\n");
1564                                 assert(0);
1565                         }
1566                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1567                 }
1568                 td->files_size = new_size;
1569         }
1570         td->files[cur_files] = f;
1571         f->fileno = cur_files;
1572
1573         /*
1574          * init function, io engine may not be loaded yet
1575          */
1576         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1577                 f->real_file_size = -1ULL;
1578
1579         f->file_name = smalloc_strdup(file_name);
1580         if (!f->file_name)
1581                 assert(0);
1582
1583         get_file_type(f);
1584
1585         switch (td->o.file_lock_mode) {
1586         case FILE_LOCK_NONE:
1587                 break;
1588         case FILE_LOCK_READWRITE:
1589                 f->rwlock = fio_rwlock_init();
1590                 break;
1591         case FILE_LOCK_EXCLUSIVE:
1592                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1593                 break;
1594         default:
1595                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1596                 assert(0);
1597         }
1598
1599         td->files_index++;
1600         if (f->filetype == FIO_TYPE_FILE)
1601                 td->nr_normal_files++;
1602
1603         set_already_allocated(file_name);
1604
1605         if (inc)
1606                 td->o.nr_files++;
1607
1608         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1609                                                         cur_files);
1610
1611         return cur_files;
1612 }
1613
1614 int add_file_exclusive(struct thread_data *td, const char *fname)
1615 {
1616         struct fio_file *f;
1617         unsigned int i;
1618
1619         for_each_file(td, f, i) {
1620                 if (!strcmp(f->file_name, fname))
1621                         return i;
1622         }
1623
1624         return add_file(td, fname, 0, 1);
1625 }
1626
1627 void get_file(struct fio_file *f)
1628 {
1629         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1630         assert(fio_file_open(f));
1631         f->references++;
1632 }
1633
1634 int put_file(struct thread_data *td, struct fio_file *f)
1635 {
1636         int f_ret = 0, ret = 0;
1637
1638         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1639
1640         if (!fio_file_open(f)) {
1641                 assert(f->fd == -1);
1642                 return 0;
1643         }
1644
1645         assert(f->references);
1646         if (--f->references)
1647                 return 0;
1648
1649         if (should_fsync(td) && td->o.fsync_on_close) {
1650                 f_ret = fsync(f->fd);
1651                 if (f_ret < 0)
1652                         f_ret = errno;
1653         }
1654
1655         if (td->io_ops->close_file)
1656                 ret = td->io_ops->close_file(td, f);
1657
1658         if (!ret)
1659                 ret = f_ret;
1660
1661         td->nr_open_files--;
1662         fio_file_clear_open(f);
1663         assert(f->fd == -1);
1664         return ret;
1665 }
1666
1667 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1668 {
1669         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1670                 return;
1671
1672         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1673                 if (ddir == DDIR_READ)
1674                         fio_rwlock_read(f->rwlock);
1675                 else
1676                         fio_rwlock_write(f->rwlock);
1677         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1678                 fio_mutex_down(f->lock);
1679
1680         td->file_locks[f->fileno] = td->o.file_lock_mode;
1681 }
1682
1683 void unlock_file(struct thread_data *td, struct fio_file *f)
1684 {
1685         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1686                 return;
1687
1688         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1689                 fio_rwlock_unlock(f->rwlock);
1690         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1691                 fio_mutex_up(f->lock);
1692
1693         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1694 }
1695
1696 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1697 {
1698         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1699                 return;
1700         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1701                 unlock_file(td, f);
1702 }
1703
1704 static int recurse_dir(struct thread_data *td, const char *dirname)
1705 {
1706         struct dirent *dir;
1707         int ret = 0;
1708         DIR *D;
1709
1710         D = opendir(dirname);
1711         if (!D) {
1712                 char buf[FIO_VERROR_SIZE];
1713
1714                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1715                 td_verror(td, errno, buf);
1716                 return 1;
1717         }
1718
1719         while ((dir = readdir(D)) != NULL) {
1720                 char full_path[PATH_MAX];
1721                 struct stat sb;
1722
1723                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1724                         continue;
1725
1726                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1727
1728                 if (lstat(full_path, &sb) == -1) {
1729                         if (errno != ENOENT) {
1730                                 td_verror(td, errno, "stat");
1731                                 ret = 1;
1732                                 break;
1733                         }
1734                 }
1735
1736                 if (S_ISREG(sb.st_mode)) {
1737                         add_file(td, full_path, 0, 1);
1738                         continue;
1739                 }
1740                 if (!S_ISDIR(sb.st_mode))
1741                         continue;
1742
1743                 ret = recurse_dir(td, full_path);
1744                 if (ret)
1745                         break;
1746         }
1747
1748         closedir(D);
1749         return ret;
1750 }
1751
1752 int add_dir_files(struct thread_data *td, const char *path)
1753 {
1754         int ret = recurse_dir(td, path);
1755
1756         if (!ret)
1757                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1758
1759         return ret;
1760 }
1761
1762 void dup_files(struct thread_data *td, struct thread_data *org)
1763 {
1764         struct fio_file *f;
1765         unsigned int i;
1766
1767         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1768
1769         if (!org->files)
1770                 return;
1771
1772         td->files = malloc(org->files_index * sizeof(f));
1773
1774         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1775                 td->file_locks = malloc(org->files_index);
1776
1777         for_each_file(org, f, i) {
1778                 struct fio_file *__f;
1779
1780                 __f = alloc_new_file(td);
1781
1782                 if (f->file_name) {
1783                         __f->file_name = smalloc_strdup(f->file_name);
1784                         if (!__f->file_name)
1785                                 assert(0);
1786
1787                         __f->filetype = f->filetype;
1788                 }
1789
1790                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1791                         __f->lock = f->lock;
1792                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1793                         __f->rwlock = f->rwlock;
1794
1795                 td->files[i] = __f;
1796         }
1797 }
1798
1799 /*
1800  * Returns the index that matches the filename, or -1 if not there
1801  */
1802 int get_fileno(struct thread_data *td, const char *fname)
1803 {
1804         struct fio_file *f;
1805         unsigned int i;
1806
1807         for_each_file(td, f, i)
1808                 if (!strcmp(f->file_name, fname))
1809                         return i;
1810
1811         return -1;
1812 }
1813
1814 /*
1815  * For log usage, where we add/open/close files automatically
1816  */
1817 void free_release_files(struct thread_data *td)
1818 {
1819         close_files(td);
1820         td->o.nr_files = 0;
1821         td->o.open_files = 0;
1822         td->files_index = 0;
1823         td->nr_normal_files = 0;
1824 }
1825
1826 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1827 {
1828         int i;
1829
1830         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1831                 f->last_pos[i] = f->file_offset;
1832                 f->last_start[i] = -1ULL;
1833         }
1834
1835         if (fio_file_axmap(f))
1836                 axmap_reset(f->io_axmap);
1837         else if (fio_file_lfsr(f))
1838                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1839 }
1840
1841 bool fio_files_done(struct thread_data *td)
1842 {
1843         struct fio_file *f;
1844         unsigned int i;
1845
1846         for_each_file(td, f, i)
1847                 if (!fio_file_done(f))
1848                         return false;
1849
1850         return true;
1851 }
1852
1853 /* free memory used in initialization phase only */
1854 void filesetup_mem_free(void)
1855 {
1856         free_already_allocated();
1857 }
1858
1859 /*
1860  * This function is for platforms which support direct I/O but not O_DIRECT.
1861  */
1862 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1863 {
1864 #ifdef FIO_OS_DIRECTIO
1865         int ret = fio_set_odirect(f);
1866
1867         if (ret) {
1868                 td_verror(td, ret, "fio_set_directio");
1869 #if defined(__sun__)
1870                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1871                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1872                 } else {
1873                         log_err("fio: the file system does not seem to support direct IO\n");
1874                 }
1875 #else
1876                 log_err("fio: the file system does not seem to support direct IO\n");
1877 #endif
1878                 return -1;
1879         }
1880
1881         return 0;
1882 #else
1883         return -1;
1884 #endif
1885 }