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 && errno != ENOSYS)
  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  * Leaves f->fd open on success, caller must close
 107  */
 108 static int extend_file(struct thread_data *td, struct fio_file *f)
 109 {
 110         int new_layout = 0, unlink_file = 0, flags;
 111         unsigned long long left;
 112         unsigned int bs;
 113         char *b = NULL;
 114
 115         if (read_only) {
 116                 log_err("fio: refusing extend of file due to read-only\n");
 117                 return 0;
 118         }
 119
 120         /*
 121          * check if we need to lay the file out complete again. fio
 122          * does that for operations involving reads, or for writes
 123          * where overwrite is set
 124          */
 125         if (td_read(td) ||
 126            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
 127             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
 128                 new_layout = 1;
 129         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
 130                 unlink_file = 1;
 131
 132         if (unlink_file || new_layout) {
 133                 int ret;
 134
 135                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
 136
 137                 ret = td_io_unlink_file(td, f);
 138                 if (ret != 0 && ret != ENOENT) {
 139                         td_verror(td, errno, "unlink");
 140                         return 1;
 141                 }
 142         }
 143
 144         flags = O_WRONLY;
 145         if (td->o.allow_create)
 146                 flags |= O_CREAT;
 147         if (new_layout)
 148                 flags |= O_TRUNC;
 149
 150 #ifdef WIN32
 151         flags |= _O_BINARY;
 152 #endif
 153
 154         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 155         f->fd = open(f->file_name, flags, 0644);
 156         if (f->fd < 0) {
 157                 int err = errno;
 158
 159                 if (err == ENOENT && !td->o.allow_create)
 160                         log_err("fio: file creation disallowed by "
 161                                         "allow_file_create=0\n");
 162                 else
 163                         td_verror(td, err, "open");
 164                 return 1;
 165         }
 166
 167         fallocate_file(td, f);
 168
 169         /*
 170          * If our jobs don't require regular files initially, we're done.
 171          */
 172         if (!new_layout)
 173                 goto done;
 174
 175         /*
 176          * The size will be -1ULL when fill_device is used, so don't truncate
 177          * or fallocate this file, just write it
 178          */
 179         if (!td->o.fill_device) {
 180                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 181                                         (unsigned long long) f->real_file_size);
 182                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 183                         if (errno != EFBIG) {
 184                                 td_verror(td, errno, "ftruncate");
 185                                 goto err;
 186                         }
 187                 }
 188         }
 189
 190         left = f->real_file_size;
 191         bs = td->o.max_bs[DDIR_WRITE];
 192         if (bs > left)
 193                 bs = left;
 194
 195         b = malloc(bs);
 196         if (!b) {
 197                 td_verror(td, errno, "malloc");
 198                 goto err;
 199         }
 200
 201         while (left && !td->terminate) {
 202                 ssize_t r;
 203
 204                 if (bs > left)
 205                         bs = left;
 206
 207                 fill_io_buffer(td, b, bs, bs);
 208
 209                 r = write(f->fd, b, bs);
 210
 211                 if (r > 0) {
 212                         left -= r;
 213                         continue;
 214                 } else {
 215                         if (r < 0) {
 216                                 int __e = errno;
 217
 218                                 if (__e == ENOSPC) {
 219                                         if (td->o.fill_device)
 220                                                 break;
 221                                         log_info("fio: ENOSPC on laying out "
 222                                                  "file, stopping\n");
 223                                         break;
 224                                 }
 225                                 td_verror(td, errno, "write");
 226                         } else
 227                                 td_verror(td, EIO, "write");
 228
 229                         break;
 230                 }
 231         }
 232
 233         if (td->terminate) {
 234                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 235                 td_io_unlink_file(td, f);
 236         } else if (td->o.create_fsync) {
 237                 if (fsync(f->fd) < 0) {
 238                         td_verror(td, errno, "fsync");
 239                         goto err;
 240                 }
 241         }
 242         if (td->o.fill_device && !td_write(td)) {
 243                 fio_file_clear_size_known(f);
 244                 if (td_io_get_file_size(td, f))
 245                         goto err;
 246                 if (f->io_size > f->real_file_size)
 247                         f->io_size = f->real_file_size;
 248         }
 249
 250         free(b);
 251 done:
 252         return 0;
 253 err:
 254         close(f->fd);
 255         f->fd = -1;
 256         if (b)
 257                 free(b);
 258         return 1;
 259 }
 260
 261 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 262 {
 263         int ret = 0, r, did_open = 0, old_runstate;
 264         unsigned long long left;
 265         unsigned int bs;
 266         char *b;
 267
 268         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 269             td_ioengine_flagged(td, FIO_NOIO))
 270                 return 0;
 271
 272         if (f->filetype == FIO_TYPE_CHAR)
 273                 return 0;
 274
 275         if (!fio_file_open(f)) {
 276                 if (td->io_ops->open_file(td, f)) {
 277                         log_err("fio: cannot pre-read, failed to open file\n");
 278                         return 1;
 279                 }
 280                 did_open = 1;
 281         }
 282
 283         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 284
 285         left = f->io_size;
 286         bs = td->o.max_bs[DDIR_READ];
 287         if (bs > left)
 288                 bs = left;
 289
 290         b = malloc(bs);
 291         if (!b) {
 292                 td_verror(td, errno, "malloc");
 293                 ret = 1;
 294                 goto error;
 295         }
 296         memset(b, 0, bs);
 297
 298         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 299                 td_verror(td, errno, "lseek");
 300                 log_err("fio: failed to lseek pre-read file\n");
 301                 ret = 1;
 302                 goto error;
 303         }
 304
 305         while (left && !td->terminate) {
 306                 if (bs > left)
 307                         bs = left;
 308
 309                 r = read(f->fd, b, bs);
 310
 311                 if (r == (int) bs) {
 312                         left -= bs;
 313                         continue;
 314                 } else {
 315                         td_verror(td, EIO, "pre_read");
 316                         break;
 317                 }
 318         }
 319
 320 error:
 321         td_restore_runstate(td, old_runstate);
 322
 323         if (did_open)
 324                 td->io_ops->close_file(td, f);
 325
 326         free(b);
 327         return ret;
 328 }
 329
 330 unsigned long long get_rand_file_size(struct thread_data *td)
 331 {
 332         unsigned long long ret, sized;
 333         uint64_t frand_max;
 334         unsigned long r;
 335
 336         frand_max = rand_max(&td->file_size_state);
 337         r = __rand(&td->file_size_state);
 338         sized = td->o.file_size_high - td->o.file_size_low;
 339         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 340         ret += td->o.file_size_low;
 341         ret -= (ret % td->o.rw_min_bs);
 342         return ret;
 343 }
 344
 345 static int file_size(struct thread_data *td, struct fio_file *f)
 346 {
 347         struct stat st;
 348
 349         if (stat(f->file_name, &st) == -1) {
 350                 td_verror(td, errno, "fstat");
 351                 return 1;
 352         }
 353
 354         f->real_file_size = st.st_size;
 355         return 0;
 356 }
 357
 358 static int bdev_size(struct thread_data *td, struct fio_file *f)
 359 {
 360         unsigned long long bytes = 0;
 361         int r;
 362
 363         if (td->io_ops->open_file(td, f)) {
 364                 log_err("fio: failed opening blockdev %s for size check\n",
 365                         f->file_name);
 366                 return 1;
 367         }
 368
 369         r = blockdev_size(f, &bytes);
 370         if (r) {
 371                 td_verror(td, r, "blockdev_size");
 372                 goto err;
 373         }
 374
 375         if (!bytes) {
 376                 log_err("%s: zero sized block device?\n", f->file_name);
 377                 goto err;
 378         }
 379
 380         f->real_file_size = bytes;
 381         td->io_ops->close_file(td, f);
 382         return 0;
 383 err:
 384         td->io_ops->close_file(td, f);
 385         return 1;
 386 }
 387
 388 static int char_size(struct thread_data *td, struct fio_file *f)
 389 {
 390 #ifdef FIO_HAVE_CHARDEV_SIZE
 391         unsigned long long bytes = 0;
 392         int r;
 393
 394         if (td->io_ops->open_file(td, f)) {
 395                 log_err("fio: failed opening chardev %s for size check\n",
 396                         f->file_name);
 397                 return 1;
 398         }
 399
 400         r = chardev_size(f, &bytes);
 401         if (r) {
 402                 td_verror(td, r, "chardev_size");
 403                 goto err;
 404         }
 405
 406         if (!bytes) {
 407                 log_err("%s: zero sized char device?\n", f->file_name);
 408                 goto err;
 409         }
 410
 411         f->real_file_size = bytes;
 412         td->io_ops->close_file(td, f);
 413         return 0;
 414 err:
 415         td->io_ops->close_file(td, f);
 416         return 1;
 417 #else
 418         f->real_file_size = -1ULL;
 419         return 0;
 420 #endif
 421 }
 422
 423 static int get_file_size(struct thread_data *td, struct fio_file *f)
 424 {
 425         int ret = 0;
 426
 427         if (fio_file_size_known(f))
 428                 return 0;
 429
 430         if (f->filetype == FIO_TYPE_FILE)
 431                 ret = file_size(td, f);
 432         else if (f->filetype == FIO_TYPE_BLOCK)
 433                 ret = bdev_size(td, f);
 434         else if (f->filetype == FIO_TYPE_CHAR)
 435                 ret = char_size(td, f);
 436         else
 437                 f->real_file_size = -1ULL;
 438
 439         /*
 440          * Leave ->real_file_size with 0 since it could be expectation
 441          * of initial setup for regular files.
 442          */
 443         if (ret)
 444                 return ret;
 445
 446         /*
 447          * If ->real_file_size is -1, a conditional for the message
 448          * "offset extends end" is always true, but it makes no sense,
 449          * so just return the same value here.
 450          */
 451         if (f->real_file_size == -1ULL) {
 452                 log_info("%s: failed to get file size of %s\n", td->o.name,
 453                                         f->file_name);
 454                 return 1;
 455         }
 456
 457         if (td->o.start_offset && f->file_offset == 0)
 458                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 459                                         f->file_name);
 460         /*
 461          * ->file_offset normally hasn't been initialized yet, so this
 462          * is basically always false.
 463          */
 464         if (f->file_offset > f->real_file_size) {
 465                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 466                                         (unsigned long long) f->file_offset,
 467                                         (unsigned long long) f->real_file_size);
 468                 return 1;
 469         }
 470
 471         fio_file_set_size_known(f);
 472         return 0;
 473 }
 474
 475 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 476                                    unsigned long long off,
 477                                    unsigned long long len)
 478 {
 479         int errval = 0, ret = 0;
 480
 481 #ifdef CONFIG_ESX
 482         return 0;
 483 #endif
 484
 485         if (len == -1ULL)
 486                 len = f->io_size;
 487         if (off == -1ULL)
 488                 off = f->file_offset;
 489
 490         if (len == -1ULL || off == -1ULL)
 491                 return 0;
 492
 493         if (td->io_ops->invalidate) {
 494                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 495                         f->file_name);
 496                 ret = td->io_ops->invalidate(td, f);
 497                 if (ret < 0)
 498                         errval = -ret;
 499         } else if (f->filetype == FIO_TYPE_FILE) {
 500                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 501                         f->file_name, off, len);
 502                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 503                 if (ret)
 504                         errval = ret;
 505         } else if (f->filetype == FIO_TYPE_BLOCK) {
 506                 int retry_count = 0;
 507
 508                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 509                 ret = blockdev_invalidate_cache(f);
 510                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 511                         /*
 512                          * Linux multipath devices reject ioctl while
 513                          * the maps are being updated. That window can
 514                          * last tens of milliseconds; we'll try up to
 515                          * a quarter of a second.
 516                          */
 517                         usleep(10000);
 518                         ret = blockdev_invalidate_cache(f);
 519                 }
 520                 if (ret < 0 && errno == EACCES && geteuid()) {
 521                         if (!root_warn) {
 522                                 log_err("fio: only root may flush block "
 523                                         "devices. Cache flush bypassed!\n");
 524                                 root_warn = 1;
 525                         }
 526                         ret = 0;
 527                 }
 528                 if (ret < 0)
 529                         errval = errno;
 530         } else if (f->filetype == FIO_TYPE_CHAR ||
 531                    f->filetype == FIO_TYPE_PIPE) {
 532                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 533                 ret = 0;
 534         }
 535
 536         /*
 537          * Cache flushing isn't a fatal condition, and we know it will
 538          * happen on some platforms where we don't have the proper
 539          * function to flush eg block device caches. So just warn and
 540          * continue on our way.
 541          */
 542         if (errval)
 543                 log_info("fio: cache invalidation of %s failed: %s\n",
 544                          f->file_name, strerror(errval));
 545
 546         return 0;
 547
 548 }
 549
 550 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 551 {
 552         if (!fio_file_open(f))
 553                 return 0;
 554
 555         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 556 }
 557
 558 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 559 {
 560         int ret = 0;
 561
 562         dprint(FD_FILE, "fd close %s\n", f->file_name);
 563
 564         remove_file_hash(f);
 565
 566         if (close(f->fd) < 0)
 567                 ret = errno;
 568
 569         f->fd = -1;
 570
 571         if (f->shadow_fd != -1) {
 572                 close(f->shadow_fd);
 573                 f->shadow_fd = -1;
 574         }
 575
 576         f->engine_pos = 0;
 577         return ret;
 578 }
 579
 580 int file_lookup_open(struct fio_file *f, int flags)
 581 {
 582         struct fio_file *__f;
 583         int from_hash;
 584
 585         __f = lookup_file_hash(f->file_name);
 586         if (__f) {
 587                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 588                 f->lock = __f->lock;
 589                 from_hash = 1;
 590         } else {
 591                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 592                 from_hash = 0;
 593         }
 594
 595 #ifdef WIN32
 596         flags |= _O_BINARY;
 597 #endif
 598
 599         f->fd = open(f->file_name, flags, 0600);
 600         return from_hash;
 601 }
 602
 603 static int file_close_shadow_fds(struct thread_data *td)
 604 {
 605         struct fio_file *f;
 606         int num_closed = 0;
 607         unsigned int i;
 608
 609         for_each_file(td, f, i) {
 610                 if (f->shadow_fd == -1)
 611                         continue;
 612
 613                 close(f->shadow_fd);
 614                 f->shadow_fd = -1;
 615                 num_closed++;
 616         }
 617
 618         return num_closed;
 619 }
 620
 621 int generic_open_file(struct thread_data *td, struct fio_file *f)
 622 {
 623         int is_std = 0;
 624         int flags = 0;
 625         int from_hash = 0;
 626
 627         dprint(FD_FILE, "fd open %s\n", f->file_name);
 628
 629         if (!strcmp(f->file_name, "-")) {
 630                 if (td_rw(td)) {
 631                         log_err("fio: can't read/write to stdin/out\n");
 632                         return 1;
 633                 }
 634                 is_std = 1;
 635
 636                 /*
 637                  * move output logging to stderr, if we are writing to stdout
 638                  */
 639                 if (td_write(td))
 640                         f_out = stderr;
 641         }
 642
 643         if (td_trim(td))
 644                 goto skip_flags;
 645         if (td->o.odirect)
 646                 flags |= OS_O_DIRECT;
 647         if (td->o.oatomic) {
 648                 if (!FIO_O_ATOMIC) {
 649                         td_verror(td, EINVAL, "OS does not support atomic IO");
 650                         return 1;
 651                 }
 652                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 653         }
 654         if (td->o.sync_io)
 655                 flags |= O_SYNC;
 656         if (td->o.create_on_open && td->o.allow_create)
 657                 flags |= O_CREAT;
 658 skip_flags:
 659         if (f->filetype != FIO_TYPE_FILE)
 660                 flags |= FIO_O_NOATIME;
 661
 662 open_again:
 663         if (td_write(td)) {
 664                 if (!read_only)
 665                         flags |= O_RDWR;
 666
 667                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 668                         flags |= O_CREAT;
 669
 670                 if (is_std)
 671                         f->fd = dup(STDOUT_FILENO);
 672                 else
 673                         from_hash = file_lookup_open(f, flags);
 674         } else if (td_read(td)) {
 675                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 676                         flags |= O_RDWR;
 677                 else
 678                         flags |= O_RDONLY;
 679
 680                 if (is_std)
 681                         f->fd = dup(STDIN_FILENO);
 682                 else
 683                         from_hash = file_lookup_open(f, flags);
 684         } else if (td_trim(td)) {
 685                 assert(!td_rw(td)); /* should have matched above */
 686                 flags |= O_RDWR;
 687                 from_hash = file_lookup_open(f, flags);
 688         }
 689
 690         if (f->fd == -1) {
 691                 char buf[FIO_VERROR_SIZE];
 692                 int __e = errno;
 693
 694                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 695                         flags &= ~FIO_O_NOATIME;
 696                         goto open_again;
 697                 }
 698                 if (__e == EMFILE && file_close_shadow_fds(td))
 699                         goto open_again;
 700
 701                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 702
 703                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 704                         log_err("fio: looks like your file system does not " \
 705                                 "support direct=1/buffered=0\n");
 706                 }
 707
 708                 td_verror(td, __e, buf);
 709                 return 1;
 710         }
 711
 712         if (!from_hash && f->fd != -1) {
 713                 if (add_file_hash(f)) {
 714                         int fio_unused ret;
 715
 716                         /*
 717                          * Stash away descriptor for later close. This is to
 718                          * work-around a "feature" on Linux, where a close of
 719                          * an fd that has been opened for write will trigger
 720                          * udev to call blkid to check partitions, fs id, etc.
 721                          * That pollutes the device cache, which can slow down
 722                          * unbuffered accesses.
 723                          */
 724                         if (f->shadow_fd == -1)
 725                                 f->shadow_fd = f->fd;
 726                         else {
 727                                 /*
 728                                  * OK to ignore, we haven't done anything
 729                                  * with it
 730                                  */
 731                                 ret = generic_close_file(td, f);
 732                         }
 733                         goto open_again;
 734                 }
 735         }
 736
 737         return 0;
 738 }
 739
 740 /*
 741  * This function i.e. get_file_size() is the default .get_file_size
 742  * implementation of majority of I/O engines.
 743  */
 744 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 745 {
 746         return get_file_size(td, f);
 747 }
 748
 749 /*
 750  * open/close all files, so that ->real_file_size gets set
 751  */
 752 static int get_file_sizes(struct thread_data *td)
 753 {
 754         struct fio_file *f;
 755         unsigned int i;
 756         int err = 0;
 757
 758         for_each_file(td, f, i) {
 759                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 760                                                                 f->file_name);
 761
 762                 if (td_io_get_file_size(td, f)) {
 763                         if (td->error != ENOENT) {
 764                                 log_err("%s\n", td->verror);
 765                                 err = 1;
 766                                 break;
 767                         }
 768                         clear_error(td);
 769                 }
 770
 771                 /*
 772                  * There are corner cases where we end up with -1 for
 773                  * ->real_file_size due to unsupported file type, etc.
 774                  * We then just set to size option value divided by number
 775                  * of files, similar to the way file ->io_size is set.
 776                  * stat(2) failure doesn't set ->real_file_size to -1.
 777                  */
 778                 if (f->real_file_size == -1ULL && td->o.size)
 779                         f->real_file_size = td->o.size / td->o.nr_files;
 780         }
 781
 782         return err;
 783 }
 784
 785 struct fio_mount {
 786         struct flist_head list;
 787         const char *base;
 788         char __base[256];
 789         unsigned int key;
 790 };
 791
 792 /*
 793  * Get free number of bytes for each file on each unique mount.
 794  */
 795 static unsigned long long get_fs_free_counts(struct thread_data *td)
 796 {
 797         struct flist_head *n, *tmp;
 798         unsigned long long ret = 0;
 799         struct fio_mount *fm;
 800         FLIST_HEAD(list);
 801         struct fio_file *f;
 802         unsigned int i;
 803
 804         for_each_file(td, f, i) {
 805                 struct stat sb;
 806                 char buf[256];
 807
 808                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 809                         if (f->real_file_size != -1ULL)
 810                                 ret += f->real_file_size;
 811                         continue;
 812                 } else if (f->filetype != FIO_TYPE_FILE)
 813                         continue;
 814
 815                 buf[255] = '\0';
 816                 strncpy(buf, f->file_name, 255);
 817
 818                 if (stat(buf, &sb) < 0) {
 819                         if (errno != ENOENT)
 820                                 break;
 821                         strcpy(buf, ".");
 822                         if (stat(buf, &sb) < 0)
 823                                 break;
 824                 }
 825
 826                 fm = NULL;
 827                 flist_for_each(n, &list) {
 828                         fm = flist_entry(n, struct fio_mount, list);
 829                         if (fm->key == sb.st_dev)
 830                                 break;
 831
 832                         fm = NULL;
 833                 }
 834
 835                 if (fm)
 836                         continue;
 837
 838                 fm = calloc(1, sizeof(*fm));
 839                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 840                 fm->base = basename(fm->__base);
 841                 fm->key = sb.st_dev;
 842                 flist_add(&fm->list, &list);
 843         }
 844
 845         flist_for_each_safe(n, tmp, &list) {
 846                 unsigned long long sz;
 847
 848                 fm = flist_entry(n, struct fio_mount, list);
 849                 flist_del(&fm->list);
 850
 851                 sz = get_fs_free_size(fm->base);
 852                 if (sz && sz != -1ULL)
 853                         ret += sz;
 854
 855                 free(fm);
 856         }
 857
 858         return ret;
 859 }
 860
 861 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 862 {
 863         struct thread_options *o = &td->o;
 864         unsigned long long align_bs;
 865         unsigned long long offset;
 866
 867         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 868                 return f->real_file_size;
 869
 870         if (o->start_offset_percent > 0) {
 871                 /*
 872                  * if offset_align is provided, set initial offset
 873                  */
 874                 if (fio_option_is_set(o, start_offset_align)) {
 875                         align_bs = o->start_offset_align;
 876                 } else {
 877                         /* else take the minimum block size */
 878                         align_bs = td_min_bs(td);
 879                 }
 880
 881                 /* calculate the raw offset */
 882                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 883                         (td->subjob_number * o->offset_increment);
 884
 885                 /*
 886                  * block align the offset at the next available boundary at
 887                  * ceiling(offset / align_bs) * align_bs
 888                  */
 889                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 890
 891         } else {
 892                 /* start_offset_percent not set */
 893                 offset = o->start_offset +
 894                                 td->subjob_number * o->offset_increment;
 895         }
 896
 897         return offset;
 898 }
 899
 900 /*
 901  * Open the files and setup files sizes, creating files if necessary.
 902  */
 903 int setup_files(struct thread_data *td)
 904 {
 905         unsigned long long total_size, extend_size;
 906         struct thread_options *o = &td->o;
 907         struct fio_file *f;
 908         unsigned int i, nr_fs_extra = 0;
 909         int err = 0, need_extend;
 910         int old_state;
 911         const unsigned int bs = td_min_bs(td);
 912         uint64_t fs = 0;
 913
 914         dprint(FD_FILE, "setup files\n");
 915
 916         old_state = td_bump_runstate(td, TD_SETTING_UP);
 917
 918         if (o->read_iolog_file)
 919                 goto done;
 920
 921         /*
 922          * Find out physical size of files or devices for this thread,
 923          * before we determine I/O size and range of our targets.
 924          * If ioengine defines a setup() method, it's responsible for
 925          * opening the files and setting f->real_file_size to indicate
 926          * the valid range for that file.
 927          */
 928         if (td->io_ops->setup)
 929                 err = td->io_ops->setup(td);
 930         else
 931                 err = get_file_sizes(td);
 932
 933         if (err)
 934                 goto err_out;
 935
 936         /*
 937          * check sizes. if the files/devices do not exist and the size
 938          * isn't passed to fio, abort.
 939          */
 940         total_size = 0;
 941         for_each_file(td, f, i) {
 942                 f->fileno = i;
 943                 if (f->real_file_size == -1ULL)
 944                         total_size = -1ULL;
 945                 else
 946                         total_size += f->real_file_size;
 947         }
 948
 949         if (o->fill_device)
 950                 td->fill_device_size = get_fs_free_counts(td);
 951
 952         /*
 953          * device/file sizes are zero and no size given, punt
 954          */
 955         if ((!total_size || total_size == -1ULL) && !o->size &&
 956             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 957             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 958                 log_err("%s: you need to specify size=\n", o->name);
 959                 td_verror(td, EINVAL, "total_file_size");
 960                 goto err_out;
 961         }
 962
 963         /*
 964          * Calculate per-file size and potential extra size for the
 965          * first files, if needed (i.e. if we don't have a fixed size).
 966          */
 967         if (!o->file_size_low && o->nr_files) {
 968                 uint64_t all_fs;
 969
 970                 fs = o->size / o->nr_files;
 971                 all_fs = fs * o->nr_files;
 972
 973                 if (all_fs < o->size)
 974                         nr_fs_extra = (o->size - all_fs) / bs;
 975         }
 976
 977         /*
 978          * now file sizes are known, so we can set ->io_size. if size= is
 979          * not given, ->io_size is just equal to ->real_file_size. if size
 980          * is given, ->io_size is size / nr_files.
 981          */
 982         extend_size = total_size = 0;
 983         need_extend = 0;
 984         for_each_file(td, f, i) {
 985                 f->file_offset = get_start_offset(td, f);
 986
 987                 /*
 988                  * Update ->io_size depending on options specified.
 989                  * ->file_size_low being 0 means filesize option isn't set.
 990                  * Non zero ->file_size_low equals ->file_size_high means
 991                  * filesize option is set in a fixed size format.
 992                  * Non zero ->file_size_low not equals ->file_size_high means
 993                  * filesize option is set in a range format.
 994                  */
 995                 if (!o->file_size_low) {
 996                         /*
 997                          * no file size or range given, file size is equal to
 998                          * total size divided by number of files. If the size
 999                          * doesn't divide nicely with the min blocksize,
1000                          * make the first files bigger.
1001                          */
1002                         f->io_size = fs;
1003                         if (nr_fs_extra) {
1004                                 nr_fs_extra--;
1005                                 f->io_size += bs;
1006                         }
1007
1008                         /*
1009                          * We normally don't come here for regular files, but
1010                          * if the result is 0 for a regular file, set it to the
1011                          * real file size. This could be size of the existing
1012                          * one if it already exists, but otherwise will be set
1013                          * to 0. A new file won't be created because
1014                          * ->io_size + ->file_offset equals ->real_file_size.
1015                          */
1016                         if (!f->io_size) {
1017                                 if (f->file_offset > f->real_file_size)
1018                                         goto err_offset;
1019                                 f->io_size = f->real_file_size - f->file_offset;
1020                                 if (!f->io_size)
1021                                         log_info("fio: file %s may be ignored\n",
1022                                                 f->file_name);
1023                         }
1024                 } else if (f->real_file_size < o->file_size_low ||
1025                            f->real_file_size > o->file_size_high) {
1026                         if (f->file_offset > o->file_size_low)
1027                                 goto err_offset;
1028                         /*
1029                          * file size given. if it's fixed, use that. if it's a
1030                          * range, generate a random size in-between.
1031                          */
1032                         if (o->file_size_low == o->file_size_high)
1033                                 f->io_size = o->file_size_low - f->file_offset;
1034                         else {
1035                                 f->io_size = get_rand_file_size(td)
1036                                                 - f->file_offset;
1037                         }
1038                 } else
1039                         f->io_size = f->real_file_size - f->file_offset;
1040
1041                 if (f->io_size == -1ULL)
1042                         total_size = -1ULL;
1043                 else {
1044                         if (o->size_percent) {
1045                                 uint64_t file_size;
1046
1047                                 file_size = f->io_size + f->file_offset;
1048                                 f->io_size = (file_size *
1049                                               o->size_percent) / 100;
1050                                 if (f->io_size > (file_size - f->file_offset))
1051                                         f->io_size = file_size - f->file_offset;
1052
1053                                 f->io_size -= (f->io_size % td_min_bs(td));
1054                         }
1055                         total_size += f->io_size;
1056                 }
1057
1058                 if (f->filetype == FIO_TYPE_FILE &&
1059                     (f->io_size + f->file_offset) > f->real_file_size &&
1060                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1061                         if (!o->create_on_open) {
1062                                 need_extend++;
1063                                 extend_size += (f->io_size + f->file_offset);
1064                                 fio_file_set_extend(f);
1065                         } else
1066                                 f->real_file_size = f->io_size + f->file_offset;
1067                 }
1068         }
1069
1070         if (td->o.block_error_hist) {
1071                 int len;
1072
1073                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1074                 f = td->files[0];
1075                 len = f->io_size / td->o.bs[DDIR_TRIM];
1076                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1077                         log_err("fio: cannot calculate block histogram with "
1078                                 "%d trim blocks, maximum %d\n",
1079                                 len, MAX_NR_BLOCK_INFOS);
1080                         td_verror(td, EINVAL, "block_error_hist");
1081                         goto err_out;
1082                 }
1083
1084                 td->ts.nr_block_infos = len;
1085                 for (i = 0; i < len; i++)
1086                         td->ts.block_infos[i] =
1087                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1088         } else
1089                 td->ts.nr_block_infos = 0;
1090
1091         if (!o->size || (total_size && o->size > total_size))
1092                 o->size = total_size;
1093
1094         if (o->size < td_min_bs(td)) {
1095                 log_err("fio: blocksize too large for data set\n");
1096                 goto err_out;
1097         }
1098
1099         /*
1100          * See if we need to extend some files, typically needed when our
1101          * target regular files don't exist yet, but our jobs require them
1102          * initially due to read I/Os.
1103          */
1104         if (need_extend) {
1105                 temp_stall_ts = 1;
1106                 if (output_format & FIO_OUTPUT_NORMAL) {
1107                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1108                                  o->name,
1109                                  need_extend > 1 ? "s" : "",
1110                                  need_extend,
1111                                  need_extend > 1 ? "s" : "",
1112                                  need_extend > 1 ? "total " : "",
1113                                  extend_size >> 20);
1114                 }
1115
1116                 for_each_file(td, f, i) {
1117                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1118
1119                         if (!fio_file_extend(f))
1120                                 continue;
1121
1122                         assert(f->filetype == FIO_TYPE_FILE);
1123                         fio_file_clear_extend(f);
1124                         if (!o->fill_device) {
1125                                 old_len = f->real_file_size;
1126                                 extend_len = f->io_size + f->file_offset -
1127                                                 old_len;
1128                         }
1129                         f->real_file_size = (f->io_size + f->file_offset);
1130                         err = extend_file(td, f);
1131                         if (err)
1132                                 break;
1133
1134                         err = __file_invalidate_cache(td, f, old_len,
1135                                                                 extend_len);
1136
1137                         /*
1138                          * Shut up static checker
1139                          */
1140                         if (f->fd != -1)
1141                                 close(f->fd);
1142
1143                         f->fd = -1;
1144                         if (err)
1145                                 break;
1146                 }
1147                 temp_stall_ts = 0;
1148         }
1149
1150         if (err)
1151                 goto err_out;
1152
1153         if (!o->zone_size)
1154                 o->zone_size = o->size;
1155
1156         /*
1157          * iolog already set the total io size, if we read back
1158          * stored entries.
1159          */
1160         if (!o->read_iolog_file) {
1161                 if (o->io_size)
1162                         td->total_io_size = o->io_size * o->loops;
1163                 else
1164                         td->total_io_size = o->size * o->loops;
1165         }
1166
1167 done:
1168         if (o->create_only)
1169                 td->done = 1;
1170
1171         td_restore_runstate(td, old_state);
1172         return 0;
1173 err_offset:
1174         log_err("%s: you need to specify valid offset=\n", o->name);
1175 err_out:
1176         td_restore_runstate(td, old_state);
1177         return 1;
1178 }
1179
1180 int pre_read_files(struct thread_data *td)
1181 {
1182         struct fio_file *f;
1183         unsigned int i;
1184
1185         dprint(FD_FILE, "pre_read files\n");
1186
1187         for_each_file(td, f, i) {
1188                 if (pre_read_file(td, f))
1189                         return -1;
1190         }
1191
1192         return 0;
1193 }
1194
1195 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1196 {
1197         unsigned int range_size, seed;
1198         unsigned long nranges;
1199         uint64_t fsize;
1200
1201         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1202         fsize = min(f->real_file_size, f->io_size);
1203
1204         nranges = (fsize + range_size - 1) / range_size;
1205
1206         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1207         if (!td->o.rand_repeatable)
1208                 seed = td->rand_seeds[4];
1209
1210         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1211                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1212         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1213                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1214         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1215                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1216
1217         return 1;
1218 }
1219
1220 static int init_rand_distribution(struct thread_data *td)
1221 {
1222         struct fio_file *f;
1223         unsigned int i;
1224         int state;
1225
1226         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1227                 return 0;
1228
1229         state = td_bump_runstate(td, TD_SETTING_UP);
1230
1231         for_each_file(td, f, i)
1232                 __init_rand_distribution(td, f);
1233
1234         td_restore_runstate(td, state);
1235
1236         return 1;
1237 }
1238
1239 /*
1240  * Check if the number of blocks exceeds the randomness capability of
1241  * the selected generator. Tausworthe is 32-bit, the others are fullly
1242  * 64-bit capable.
1243  */
1244 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1245                                  uint64_t blocks)
1246 {
1247         if (blocks <= FRAND32_MAX)
1248                 return 0;
1249         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1250                 return 0;
1251
1252         /*
1253          * If the user hasn't specified a random generator, switch
1254          * to tausworthe64 with informational warning. If the user did
1255          * specify one, just warn.
1256          */
1257         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1258                         f->file_name);
1259
1260         if (!fio_option_is_set(&td->o, random_generator)) {
1261                 log_info("fio: Switching to tausworthe64. Use the "
1262                          "random_generator= option to get rid of this "
1263                          "warning.\n");
1264                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1265                 return 0;
1266         }
1267
1268         /*
1269          * Just make this information to avoid breaking scripts.
1270          */
1271         log_info("fio: Use the random_generator= option to switch to lfsr or "
1272                          "tausworthe64.\n");
1273         return 0;
1274 }
1275
1276 int init_random_map(struct thread_data *td)
1277 {
1278         unsigned long long blocks;
1279         struct fio_file *f;
1280         unsigned int i;
1281
1282         if (init_rand_distribution(td))
1283                 return 0;
1284         if (!td_random(td))
1285                 return 0;
1286
1287         for_each_file(td, f, i) {
1288                 uint64_t fsize = min(f->real_file_size, f->io_size);
1289
1290                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1291
1292                 if (check_rand_gen_limits(td, f, blocks))
1293                         return 1;
1294
1295                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1296                         unsigned long seed;
1297
1298                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1299
1300                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1301                                 fio_file_set_lfsr(f);
1302                                 continue;
1303                         }
1304                 } else if (!td->o.norandommap) {
1305                         f->io_axmap = axmap_new(blocks);
1306                         if (f->io_axmap) {
1307                                 fio_file_set_axmap(f);
1308                                 continue;
1309                         }
1310                 } else if (td->o.norandommap)
1311                         continue;
1312
1313                 if (!td->o.softrandommap) {
1314                         log_err("fio: failed allocating random map. If running"
1315                                 " a large number of jobs, try the 'norandommap'"
1316                                 " option or set 'softrandommap'. Or give"
1317                                 " a larger --alloc-size to fio.\n");
1318                         return 1;
1319                 }
1320
1321                 log_info("fio: file %s failed allocating random map. Running "
1322                          "job without.\n", f->file_name);
1323         }
1324
1325         return 0;
1326 }
1327
1328 void close_files(struct thread_data *td)
1329 {
1330         struct fio_file *f;
1331         unsigned int i;
1332
1333         for_each_file(td, f, i) {
1334                 if (fio_file_open(f))
1335                         td_io_close_file(td, f);
1336         }
1337 }
1338
1339 void close_and_free_files(struct thread_data *td)
1340 {
1341         struct fio_file *f;
1342         unsigned int i;
1343         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1344
1345         dprint(FD_FILE, "close files\n");
1346
1347         for_each_file(td, f, i) {
1348                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1349                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1350                         td_io_unlink_file(td, f);
1351                 }
1352
1353                 if (fio_file_open(f))
1354                         td_io_close_file(td, f);
1355
1356                 remove_file_hash(f);
1357
1358                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1359                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1360                         td_io_unlink_file(td, f);
1361                 }
1362
1363                 if (use_free)
1364                         free(f->file_name);
1365                 else
1366                         sfree(f->file_name);
1367                 f->file_name = NULL;
1368                 if (fio_file_axmap(f)) {
1369                         axmap_free(f->io_axmap);
1370                         f->io_axmap = NULL;
1371                 }
1372                 if (use_free)
1373                         free(f);
1374                 else
1375                         sfree(f);
1376         }
1377
1378         td->o.filename = NULL;
1379         free(td->files);
1380         free(td->file_locks);
1381         td->files_index = 0;
1382         td->files = NULL;
1383         td->file_locks = NULL;
1384         td->o.file_lock_mode = FILE_LOCK_NONE;
1385         td->o.nr_files = 0;
1386 }
1387
1388 static void get_file_type(struct fio_file *f)
1389 {
1390         struct stat sb;
1391
1392         if (!strcmp(f->file_name, "-"))
1393                 f->filetype = FIO_TYPE_PIPE;
1394         else
1395                 f->filetype = FIO_TYPE_FILE;
1396
1397 #ifdef WIN32
1398         /* \\.\ is the device namespace in Windows, where every file is
1399          * a block device */
1400         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1401                 f->filetype = FIO_TYPE_BLOCK;
1402 #endif
1403
1404         if (!stat(f->file_name, &sb)) {
1405                 if (S_ISBLK(sb.st_mode))
1406                         f->filetype = FIO_TYPE_BLOCK;
1407                 else if (S_ISCHR(sb.st_mode))
1408                         f->filetype = FIO_TYPE_CHAR;
1409                 else if (S_ISFIFO(sb.st_mode))
1410                         f->filetype = FIO_TYPE_PIPE;
1411         }
1412 }
1413
1414 static bool __is_already_allocated(const char *fname, bool set)
1415 {
1416         struct flist_head *entry;
1417         bool ret;
1418
1419         ret = file_bloom_exists(fname, set);
1420         if (!ret)
1421                 return ret;
1422
1423         flist_for_each(entry, &filename_list) {
1424                 struct file_name *fn;
1425
1426                 fn = flist_entry(entry, struct file_name, list);
1427
1428                 if (!strcmp(fn->filename, fname))
1429                         return true;
1430         }
1431
1432         return false;
1433 }
1434
1435 static bool is_already_allocated(const char *fname)
1436 {
1437         bool ret;
1438
1439         fio_file_hash_lock();
1440         ret = __is_already_allocated(fname, false);
1441         fio_file_hash_unlock();
1442
1443         return ret;
1444 }
1445
1446 static void set_already_allocated(const char *fname)
1447 {
1448         struct file_name *fn;
1449
1450         fn = malloc(sizeof(struct file_name));
1451         fn->filename = strdup(fname);
1452
1453         fio_file_hash_lock();
1454         if (!__is_already_allocated(fname, true)) {
1455                 flist_add_tail(&fn->list, &filename_list);
1456                 fn = NULL;
1457         }
1458         fio_file_hash_unlock();
1459
1460         if (fn) {
1461                 free(fn->filename);
1462                 free(fn);
1463         }
1464 }
1465
1466 static void free_already_allocated(void)
1467 {
1468         struct flist_head *entry, *tmp;
1469         struct file_name *fn;
1470
1471         if (flist_empty(&filename_list))
1472                 return;
1473
1474         fio_file_hash_lock();
1475         flist_for_each_safe(entry, tmp, &filename_list) {
1476                 fn = flist_entry(entry, struct file_name, list);
1477                 free(fn->filename);
1478                 flist_del(&fn->list);
1479                 free(fn);
1480         }
1481
1482         fio_file_hash_unlock();
1483 }
1484
1485 static struct fio_file *alloc_new_file(struct thread_data *td)
1486 {
1487         struct fio_file *f;
1488
1489         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1490                 f = calloc(1, sizeof(*f));
1491         else
1492                 f = smalloc(sizeof(*f));
1493         if (!f) {
1494                 assert(0);
1495                 return NULL;
1496         }
1497
1498         f->fd = -1;
1499         f->shadow_fd = -1;
1500         fio_file_reset(td, f);
1501         return f;
1502 }
1503
1504 bool exists_and_not_regfile(const char *filename)
1505 {
1506         struct stat sb;
1507
1508         if (lstat(filename, &sb) == -1)
1509                 return false;
1510
1511 #ifndef WIN32 /* NOT Windows */
1512         if (S_ISREG(sb.st_mode))
1513                 return false;
1514 #else
1515         /* \\.\ is the device namespace in Windows, where every file
1516          * is a device node */
1517         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1518                 return false;
1519 #endif
1520
1521         return true;
1522 }
1523
1524 static int create_work_dirs(struct thread_data *td, const char *fname)
1525 {
1526         char path[PATH_MAX];
1527         char *start, *end;
1528
1529         if (td->o.directory) {
1530                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1531                          FIO_OS_PATH_SEPARATOR, fname);
1532                 start = strstr(path, fname);
1533         } else {
1534                 snprintf(path, PATH_MAX, "%s", fname);
1535                 start = path;
1536         }
1537
1538         end = start;
1539         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1540                 if (end == start)
1541                         break;
1542                 *end = '\0';
1543                 errno = 0;
1544 #ifdef CONFIG_HAVE_MKDIR_TWO
1545                 if (mkdir(path, 0600) && errno != EEXIST) {
1546 #else
1547                 if (mkdir(path) && errno != EEXIST) {
1548 #endif
1549                         log_err("fio: failed to create dir (%s): %d\n",
1550                                 start, errno);
1551                         return 1;
1552                 }
1553                 *end = FIO_OS_PATH_SEPARATOR;
1554                 end++;
1555         }
1556         td->flags |= TD_F_DIRS_CREATED;
1557         return 0;
1558 }
1559
1560 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1561 {
1562         int cur_files = td->files_index;
1563         char file_name[PATH_MAX];
1564         struct fio_file *f;
1565         int len = 0;
1566
1567         dprint(FD_FILE, "add file %s\n", fname);
1568
1569         if (td->o.directory)
1570                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1571                                         td->o.unique_filename);
1572
1573         sprintf(file_name + len, "%s", fname);
1574
1575         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1576             !(td->flags & TD_F_DIRS_CREATED) &&
1577             create_work_dirs(td, fname))
1578                 return 1;
1579
1580         /* clean cloned siblings using existing files */
1581         if (numjob && is_already_allocated(file_name) &&
1582             !exists_and_not_regfile(fname))
1583                 return 0;
1584
1585         f = alloc_new_file(td);
1586
1587         if (td->files_size <= td->files_index) {
1588                 unsigned int new_size = td->o.nr_files + 1;
1589
1590                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1591
1592                 td->files = realloc(td->files, new_size * sizeof(f));
1593                 if (td->files == NULL) {
1594                         log_err("fio: realloc OOM\n");
1595                         assert(0);
1596                 }
1597                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1598                         td->file_locks = realloc(td->file_locks, new_size);
1599                         if (!td->file_locks) {
1600                                 log_err("fio: realloc OOM\n");
1601                                 assert(0);
1602                         }
1603                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1604                 }
1605                 td->files_size = new_size;
1606         }
1607         td->files[cur_files] = f;
1608         f->fileno = cur_files;
1609
1610         /*
1611          * init function, io engine may not be loaded yet
1612          */
1613         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1614                 f->real_file_size = -1ULL;
1615
1616         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1617                 f->file_name = strdup(file_name);
1618         else
1619                 f->file_name = smalloc_strdup(file_name);
1620         if (!f->file_name)
1621                 assert(0);
1622
1623         get_file_type(f);
1624
1625         switch (td->o.file_lock_mode) {
1626         case FILE_LOCK_NONE:
1627                 break;
1628         case FILE_LOCK_READWRITE:
1629                 f->rwlock = fio_rwlock_init();
1630                 break;
1631         case FILE_LOCK_EXCLUSIVE:
1632                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1633                 break;
1634         default:
1635                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1636                 assert(0);
1637         }
1638
1639         td->files_index++;
1640         if (f->filetype == FIO_TYPE_FILE)
1641                 td->nr_normal_files++;
1642
1643         if (td->o.numjobs > 1)
1644                 set_already_allocated(file_name);
1645
1646         if (inc)
1647                 td->o.nr_files++;
1648
1649         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1650                                                         cur_files);
1651
1652         return cur_files;
1653 }
1654
1655 int add_file_exclusive(struct thread_data *td, const char *fname)
1656 {
1657         struct fio_file *f;
1658         unsigned int i;
1659
1660         for_each_file(td, f, i) {
1661                 if (!strcmp(f->file_name, fname))
1662                         return i;
1663         }
1664
1665         return add_file(td, fname, 0, 1);
1666 }
1667
1668 void get_file(struct fio_file *f)
1669 {
1670         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1671         assert(fio_file_open(f));
1672         f->references++;
1673 }
1674
1675 int put_file(struct thread_data *td, struct fio_file *f)
1676 {
1677         int f_ret = 0, ret = 0;
1678
1679         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1680
1681         if (!fio_file_open(f)) {
1682                 assert(f->fd == -1);
1683                 return 0;
1684         }
1685
1686         assert(f->references);
1687         if (--f->references)
1688                 return 0;
1689
1690         if (should_fsync(td) && td->o.fsync_on_close) {
1691                 f_ret = fsync(f->fd);
1692                 if (f_ret < 0)
1693                         f_ret = errno;
1694         }
1695
1696         if (td->io_ops->close_file)
1697                 ret = td->io_ops->close_file(td, f);
1698
1699         if (!ret)
1700                 ret = f_ret;
1701
1702         td->nr_open_files--;
1703         fio_file_clear_open(f);
1704         assert(f->fd == -1);
1705         return ret;
1706 }
1707
1708 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1709 {
1710         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1711                 return;
1712
1713         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1714                 if (ddir == DDIR_READ)
1715                         fio_rwlock_read(f->rwlock);
1716                 else
1717                         fio_rwlock_write(f->rwlock);
1718         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1719                 fio_mutex_down(f->lock);
1720
1721         td->file_locks[f->fileno] = td->o.file_lock_mode;
1722 }
1723
1724 void unlock_file(struct thread_data *td, struct fio_file *f)
1725 {
1726         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1727                 return;
1728
1729         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1730                 fio_rwlock_unlock(f->rwlock);
1731         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1732                 fio_mutex_up(f->lock);
1733
1734         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1735 }
1736
1737 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1738 {
1739         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1740                 return;
1741         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1742                 unlock_file(td, f);
1743 }
1744
1745 static int recurse_dir(struct thread_data *td, const char *dirname)
1746 {
1747         struct dirent *dir;
1748         int ret = 0;
1749         DIR *D;
1750
1751         D = opendir(dirname);
1752         if (!D) {
1753                 char buf[FIO_VERROR_SIZE];
1754
1755                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1756                 td_verror(td, errno, buf);
1757                 return 1;
1758         }
1759
1760         while ((dir = readdir(D)) != NULL) {
1761                 char full_path[PATH_MAX];
1762                 struct stat sb;
1763
1764                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1765                         continue;
1766
1767                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1768
1769                 if (lstat(full_path, &sb) == -1) {
1770                         if (errno != ENOENT) {
1771                                 td_verror(td, errno, "stat");
1772                                 ret = 1;
1773                                 break;
1774                         }
1775                 }
1776
1777                 if (S_ISREG(sb.st_mode)) {
1778                         add_file(td, full_path, 0, 1);
1779                         continue;
1780                 }
1781                 if (!S_ISDIR(sb.st_mode))
1782                         continue;
1783
1784                 ret = recurse_dir(td, full_path);
1785                 if (ret)
1786                         break;
1787         }
1788
1789         closedir(D);
1790         return ret;
1791 }
1792
1793 int add_dir_files(struct thread_data *td, const char *path)
1794 {
1795         int ret = recurse_dir(td, path);
1796
1797         if (!ret)
1798                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1799
1800         return ret;
1801 }
1802
1803 void dup_files(struct thread_data *td, struct thread_data *org)
1804 {
1805         struct fio_file *f;
1806         unsigned int i;
1807
1808         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1809
1810         if (!org->files)
1811                 return;
1812
1813         td->files = malloc(org->files_index * sizeof(f));
1814
1815         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1816                 td->file_locks = malloc(org->files_index);
1817
1818         for_each_file(org, f, i) {
1819                 struct fio_file *__f;
1820
1821                 __f = alloc_new_file(td);
1822
1823                 if (f->file_name) {
1824                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1825                                 __f->file_name = strdup(f->file_name);
1826                         else
1827                                 __f->file_name = smalloc_strdup(f->file_name);
1828                         if (!__f->file_name)
1829                                 assert(0);
1830
1831                         __f->filetype = f->filetype;
1832                 }
1833
1834                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1835                         __f->lock = f->lock;
1836                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1837                         __f->rwlock = f->rwlock;
1838
1839                 td->files[i] = __f;
1840         }
1841 }
1842
1843 /*
1844  * Returns the index that matches the filename, or -1 if not there
1845  */
1846 int get_fileno(struct thread_data *td, const char *fname)
1847 {
1848         struct fio_file *f;
1849         unsigned int i;
1850
1851         for_each_file(td, f, i)
1852                 if (!strcmp(f->file_name, fname))
1853                         return i;
1854
1855         return -1;
1856 }
1857
1858 /*
1859  * For log usage, where we add/open/close files automatically
1860  */
1861 void free_release_files(struct thread_data *td)
1862 {
1863         close_files(td);
1864         td->o.nr_files = 0;
1865         td->o.open_files = 0;
1866         td->files_index = 0;
1867         td->nr_normal_files = 0;
1868 }
1869
1870 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1871 {
1872         int i;
1873
1874         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1875                 f->last_pos[i] = f->file_offset;
1876                 f->last_start[i] = -1ULL;
1877         }
1878
1879         if (fio_file_axmap(f))
1880                 axmap_reset(f->io_axmap);
1881         else if (fio_file_lfsr(f))
1882                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1883 }
1884
1885 bool fio_files_done(struct thread_data *td)
1886 {
1887         struct fio_file *f;
1888         unsigned int i;
1889
1890         for_each_file(td, f, i)
1891                 if (!fio_file_done(f))
1892                         return false;
1893
1894         return true;
1895 }
1896
1897 /* free memory used in initialization phase only */
1898 void filesetup_mem_free(void)
1899 {
1900         free_already_allocated();
1901 }
1902
1903 /*
1904  * This function is for platforms which support direct I/O but not O_DIRECT.
1905  */
1906 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1907 {
1908 #ifdef FIO_OS_DIRECTIO
1909         int ret = fio_set_odirect(f);
1910
1911         if (ret) {
1912                 td_verror(td, ret, "fio_set_directio");
1913 #if defined(__sun__)
1914                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1915                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1916                 } else {
1917                         log_err("fio: the file system does not seem to support direct IO\n");
1918                 }
1919 #else
1920                 log_err("fio: the file system does not seem to support direct IO\n");
1921 #endif
1922                 return -1;
1923         }
1924
1925         return 0;
1926 #else
1927         log_err("fio: direct IO is not supported on this host operating system\n");
1928         return -1;
1929 #endif
1930 }