filesetup.c

   1 #include <unistd.h>
   2 #include <fcntl.h>
   3 #include <string.h>
   4 #include <assert.h>
   5 #include <dirent.h>
   6 #include <libgen.h>
   7 #include <sys/stat.h>
   8 #include <sys/mman.h>
   9 #include <sys/types.h>
  10
  11 #include "fio.h"
  12 #include "smalloc.h"
  13 #include "filehash.h"
  14 #include "options.h"
  15 #include "os/os.h"
  16 #include "hash.h"
  17 #include "lib/axmap.h"
  18
  19 #ifdef CONFIG_LINUX_FALLOCATE
  20 #include <linux/falloc.h>
  21 #endif
  22
  23 static int root_warn;
  24
  25 static FLIST_HEAD(filename_list);
  26
  27 /*
  28  * List entry for filename_list
  29  */
  30 struct file_name {
  31         struct flist_head list;
  32         char *filename;
  33 };
  34
  35 static inline void clear_error(struct thread_data *td)
  36 {
  37         td->error = 0;
  38         td->verror[0] = '\0';
  39 }
  40
  41 static int native_fallocate(struct thread_data *td, struct fio_file *f)
  42 {
  43         bool success;
  44
  45         success = fio_fallocate(f, 0, f->real_file_size);
  46         dprint(FD_FILE, "native fallocate of file %s size %llu was "
  47                         "%ssuccessful\n", f->file_name,
  48                         (unsigned long long) f->real_file_size,
  49                         !success ? "un": "");
  50
  51         if (success)
  52                 return false;
  53
  54         if (errno == ENOSYS)
  55                 dprint(FD_FILE, "native fallocate is not implemented\n");
  56
  57         return true;
  58 }
  59
  60 static void fallocate_file(struct thread_data *td, struct fio_file *f)
  61 {
  62         if (td->o.fill_device)
  63                 return;
  64
  65         switch (td->o.fallocate_mode) {
  66         case FIO_FALLOCATE_NATIVE:
  67                 native_fallocate(td, f);
  68                 break;
  69         case FIO_FALLOCATE_NONE:
  70                 break;
  71 #ifdef CONFIG_POSIX_FALLOCATE
  72         case FIO_FALLOCATE_POSIX: {
  73                 int r;
  74
  75                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  76                                  f->file_name,
  77                                  (unsigned long long) f->real_file_size);
  78
  79                 r = posix_fallocate(f->fd, 0, f->real_file_size);
  80                 if (r > 0)
  81                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
  82                 break;
  83                 }
  84 #endif /* CONFIG_POSIX_FALLOCATE */
  85 #ifdef CONFIG_LINUX_FALLOCATE
  86         case FIO_FALLOCATE_KEEP_SIZE: {
  87                 int r;
  88
  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                 }
  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
 151 #ifdef WIN32
 152         flags |= _O_BINARY;
 153 #endif
 154
 155         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 156         f->fd = open(f->file_name, flags, 0644);
 157         if (f->fd < 0) {
 158                 int err = errno;
 159
 160                 if (err == ENOENT && !td->o.allow_create)
 161                         log_err("fio: file creation disallowed by "
 162                                         "allow_file_create=0\n");
 163                 else
 164                         td_verror(td, err, "open");
 165                 return 1;
 166         }
 167
 168         fallocate_file(td, f);
 169
 170         /*
 171          * If our jobs don't require regular files initially, we're done.
 172          */
 173         if (!new_layout)
 174                 goto done;
 175
 176         /*
 177          * The size will be -1ULL when fill_device is used, so don't truncate
 178          * or fallocate this file, just write it
 179          */
 180         if (!td->o.fill_device) {
 181                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 182                                         (unsigned long long) f->real_file_size);
 183                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 184                         if (errno != EFBIG) {
 185                                 td_verror(td, errno, "ftruncate");
 186                                 goto err;
 187                         }
 188                 }
 189         }
 190
 191         left = f->real_file_size;
 192         bs = td->o.max_bs[DDIR_WRITE];
 193         if (bs > left)
 194                 bs = left;
 195
 196         b = malloc(bs);
 197         if (!b) {
 198                 td_verror(td, errno, "malloc");
 199                 goto err;
 200         }
 201
 202         while (left && !td->terminate) {
 203                 ssize_t r;
 204
 205                 if (bs > left)
 206                         bs = left;
 207
 208                 fill_io_buffer(td, b, bs, bs);
 209
 210                 r = write(f->fd, b, bs);
 211
 212                 if (r > 0) {
 213                         left -= r;
 214                         continue;
 215                 } else {
 216                         if (r < 0) {
 217                                 int __e = errno;
 218
 219                                 if (__e == ENOSPC) {
 220                                         if (td->o.fill_device)
 221                                                 break;
 222                                         log_info("fio: ENOSPC on laying out "
 223                                                  "file, stopping\n");
 224                                         break;
 225                                 }
 226                                 td_verror(td, errno, "write");
 227                         } else
 228                                 td_verror(td, EIO, "write");
 229
 230                         break;
 231                 }
 232         }
 233
 234         if (td->terminate) {
 235                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 236                 td_io_unlink_file(td, f);
 237         } else if (td->o.create_fsync) {
 238                 if (fsync(f->fd) < 0) {
 239                         td_verror(td, errno, "fsync");
 240                         goto err;
 241                 }
 242         }
 243         if (td->o.fill_device && !td_write(td)) {
 244                 fio_file_clear_size_known(f);
 245                 if (td_io_get_file_size(td, f))
 246                         goto err;
 247                 if (f->io_size > f->real_file_size)
 248                         f->io_size = f->real_file_size;
 249         }
 250
 251         free(b);
 252 done:
 253         return 0;
 254 err:
 255         close(f->fd);
 256         f->fd = -1;
 257         if (b)
 258                 free(b);
 259         return 1;
 260 }
 261
 262 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 263 {
 264         int r, did_open = 0, old_runstate;
 265         unsigned long long left;
 266         unsigned int bs;
 267         bool ret = true;
 268         char *b;
 269
 270         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 271             td_ioengine_flagged(td, FIO_NOIO))
 272                 return true;
 273
 274         if (f->filetype == FIO_TYPE_CHAR)
 275                 return true;
 276
 277         if (!fio_file_open(f)) {
 278                 if (td->io_ops->open_file(td, f)) {
 279                         log_err("fio: cannot pre-read, failed to open file\n");
 280                         return false;
 281                 }
 282                 did_open = 1;
 283         }
 284
 285         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 286
 287         left = f->io_size;
 288         bs = td->o.max_bs[DDIR_READ];
 289         if (bs > left)
 290                 bs = left;
 291
 292         b = malloc(bs);
 293         if (!b) {
 294                 td_verror(td, errno, "malloc");
 295                 ret = false;
 296                 goto error;
 297         }
 298         memset(b, 0, bs);
 299
 300         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 301                 td_verror(td, errno, "lseek");
 302                 log_err("fio: failed to lseek pre-read file\n");
 303                 ret = false;
 304                 goto error;
 305         }
 306
 307         while (left && !td->terminate) {
 308                 if (bs > left)
 309                         bs = left;
 310
 311                 r = read(f->fd, b, bs);
 312
 313                 if (r == (int) bs) {
 314                         left -= bs;
 315                         continue;
 316                 } else {
 317                         td_verror(td, EIO, "pre_read");
 318                         break;
 319                 }
 320         }
 321
 322 error:
 323         td_restore_runstate(td, old_runstate);
 324
 325         if (did_open)
 326                 td->io_ops->close_file(td, f);
 327
 328         free(b);
 329         return ret;
 330 }
 331
 332 unsigned long long get_rand_file_size(struct thread_data *td)
 333 {
 334         unsigned long long ret, sized;
 335         uint64_t frand_max;
 336         unsigned long r;
 337
 338         frand_max = rand_max(&td->file_size_state);
 339         r = __rand(&td->file_size_state);
 340         sized = td->o.file_size_high - td->o.file_size_low;
 341         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 342         ret += td->o.file_size_low;
 343         ret -= (ret % td->o.rw_min_bs);
 344         return ret;
 345 }
 346
 347 static int file_size(struct thread_data *td, struct fio_file *f)
 348 {
 349         struct stat st;
 350
 351         if (stat(f->file_name, &st) == -1) {
 352                 td_verror(td, errno, "fstat");
 353                 return 1;
 354         }
 355
 356         f->real_file_size = st.st_size;
 357         return 0;
 358 }
 359
 360 static int bdev_size(struct thread_data *td, struct fio_file *f)
 361 {
 362         unsigned long long bytes = 0;
 363         int r;
 364
 365         if (td->io_ops->open_file(td, f)) {
 366                 log_err("fio: failed opening blockdev %s for size check\n",
 367                         f->file_name);
 368                 return 1;
 369         }
 370
 371         r = blockdev_size(f, &bytes);
 372         if (r) {
 373                 td_verror(td, r, "blockdev_size");
 374                 goto err;
 375         }
 376
 377         if (!bytes) {
 378                 log_err("%s: zero sized block device?\n", f->file_name);
 379                 goto err;
 380         }
 381
 382         f->real_file_size = bytes;
 383         td->io_ops->close_file(td, f);
 384         return 0;
 385 err:
 386         td->io_ops->close_file(td, f);
 387         return 1;
 388 }
 389
 390 static int char_size(struct thread_data *td, struct fio_file *f)
 391 {
 392 #ifdef FIO_HAVE_CHARDEV_SIZE
 393         unsigned long long bytes = 0;
 394         int r;
 395
 396         if (td->io_ops->open_file(td, f)) {
 397                 log_err("fio: failed opening chardev %s for size check\n",
 398                         f->file_name);
 399                 return 1;
 400         }
 401
 402         r = chardev_size(f, &bytes);
 403         if (r) {
 404                 td_verror(td, r, "chardev_size");
 405                 goto err;
 406         }
 407
 408         if (!bytes) {
 409                 log_err("%s: zero sized char device?\n", f->file_name);
 410                 goto err;
 411         }
 412
 413         f->real_file_size = bytes;
 414         td->io_ops->close_file(td, f);
 415         return 0;
 416 err:
 417         td->io_ops->close_file(td, f);
 418         return 1;
 419 #else
 420         f->real_file_size = -1ULL;
 421         return 0;
 422 #endif
 423 }
 424
 425 static int get_file_size(struct thread_data *td, struct fio_file *f)
 426 {
 427         int ret = 0;
 428
 429         if (fio_file_size_known(f))
 430                 return 0;
 431
 432         if (f->filetype == FIO_TYPE_FILE)
 433                 ret = file_size(td, f);
 434         else if (f->filetype == FIO_TYPE_BLOCK)
 435                 ret = bdev_size(td, f);
 436         else if (f->filetype == FIO_TYPE_CHAR)
 437                 ret = char_size(td, f);
 438         else
 439                 f->real_file_size = -1ULL;
 440
 441         /*
 442          * Leave ->real_file_size with 0 since it could be expectation
 443          * of initial setup for regular files.
 444          */
 445         if (ret)
 446                 return ret;
 447
 448         /*
 449          * If ->real_file_size is -1, a conditional for the message
 450          * "offset extends end" is always true, but it makes no sense,
 451          * so just return the same value here.
 452          */
 453         if (f->real_file_size == -1ULL) {
 454                 log_info("%s: failed to get file size of %s\n", td->o.name,
 455                                         f->file_name);
 456                 return 1;
 457         }
 458
 459         if (td->o.start_offset && f->file_offset == 0)
 460                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 461                                         f->file_name);
 462         /*
 463          * ->file_offset normally hasn't been initialized yet, so this
 464          * is basically always false.
 465          */
 466         if (f->file_offset > f->real_file_size) {
 467                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 468                                         (unsigned long long) f->file_offset,
 469                                         (unsigned long long) f->real_file_size);
 470                 return 1;
 471         }
 472
 473         fio_file_set_size_known(f);
 474         return 0;
 475 }
 476
 477 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 478                                    unsigned long long off,
 479                                    unsigned long long len)
 480 {
 481         int errval = 0, ret = 0;
 482
 483 #ifdef CONFIG_ESX
 484         return 0;
 485 #endif
 486
 487         if (len == -1ULL)
 488                 len = f->io_size;
 489         if (off == -1ULL)
 490                 off = f->file_offset;
 491
 492         if (len == -1ULL || off == -1ULL)
 493                 return 0;
 494
 495         if (td->io_ops->invalidate) {
 496                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 497                         f->file_name);
 498                 ret = td->io_ops->invalidate(td, f);
 499                 if (ret < 0)
 500                         errval = -ret;
 501         } else if (f->filetype == FIO_TYPE_FILE) {
 502                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 503                         f->file_name, off, len);
 504                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 505                 if (ret)
 506                         errval = ret;
 507         } else if (f->filetype == FIO_TYPE_BLOCK) {
 508                 int retry_count = 0;
 509
 510                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 511                 ret = blockdev_invalidate_cache(f);
 512                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 513                         /*
 514                          * Linux multipath devices reject ioctl while
 515                          * the maps are being updated. That window can
 516                          * last tens of milliseconds; we'll try up to
 517                          * a quarter of a second.
 518                          */
 519                         usleep(10000);
 520                         ret = blockdev_invalidate_cache(f);
 521                 }
 522                 if (ret < 0 && errno == EACCES && geteuid()) {
 523                         if (!root_warn) {
 524                                 log_err("fio: only root may flush block "
 525                                         "devices. Cache flush bypassed!\n");
 526                                 root_warn = 1;
 527                         }
 528                         ret = 0;
 529                 }
 530                 if (ret < 0)
 531                         errval = errno;
 532         } else if (f->filetype == FIO_TYPE_CHAR ||
 533                    f->filetype == FIO_TYPE_PIPE) {
 534                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 535                 ret = 0;
 536         }
 537
 538         /*
 539          * Cache flushing isn't a fatal condition, and we know it will
 540          * happen on some platforms where we don't have the proper
 541          * function to flush eg block device caches. So just warn and
 542          * continue on our way.
 543          */
 544         if (errval)
 545                 log_info("fio: cache invalidation of %s failed: %s\n",
 546                          f->file_name, strerror(errval));
 547
 548         return 0;
 549
 550 }
 551
 552 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 553 {
 554         if (!fio_file_open(f))
 555                 return 0;
 556
 557         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 558 }
 559
 560 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 561 {
 562         int ret = 0;
 563
 564         dprint(FD_FILE, "fd close %s\n", f->file_name);
 565
 566         remove_file_hash(f);
 567
 568         if (close(f->fd) < 0)
 569                 ret = errno;
 570
 571         f->fd = -1;
 572
 573         if (f->shadow_fd != -1) {
 574                 close(f->shadow_fd);
 575                 f->shadow_fd = -1;
 576         }
 577
 578         f->engine_pos = 0;
 579         return ret;
 580 }
 581
 582 int file_lookup_open(struct fio_file *f, int flags)
 583 {
 584         struct fio_file *__f;
 585         int from_hash;
 586
 587         __f = lookup_file_hash(f->file_name);
 588         if (__f) {
 589                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 590                 f->lock = __f->lock;
 591                 from_hash = 1;
 592         } else {
 593                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 594                 from_hash = 0;
 595         }
 596
 597 #ifdef WIN32
 598         flags |= _O_BINARY;
 599 #endif
 600
 601         f->fd = open(f->file_name, flags, 0600);
 602         return from_hash;
 603 }
 604
 605 static int file_close_shadow_fds(struct thread_data *td)
 606 {
 607         struct fio_file *f;
 608         int num_closed = 0;
 609         unsigned int i;
 610
 611         for_each_file(td, f, i) {
 612                 if (f->shadow_fd == -1)
 613                         continue;
 614
 615                 close(f->shadow_fd);
 616                 f->shadow_fd = -1;
 617                 num_closed++;
 618         }
 619
 620         return num_closed;
 621 }
 622
 623 int generic_open_file(struct thread_data *td, struct fio_file *f)
 624 {
 625         int is_std = 0;
 626         int flags = 0;
 627         int from_hash = 0;
 628
 629         dprint(FD_FILE, "fd open %s\n", f->file_name);
 630
 631         if (!strcmp(f->file_name, "-")) {
 632                 if (td_rw(td)) {
 633                         log_err("fio: can't read/write to stdin/out\n");
 634                         return 1;
 635                 }
 636                 is_std = 1;
 637
 638                 /*
 639                  * move output logging to stderr, if we are writing to stdout
 640                  */
 641                 if (td_write(td))
 642                         f_out = stderr;
 643         }
 644
 645         if (td_trim(td))
 646                 goto skip_flags;
 647         if (td->o.odirect)
 648                 flags |= OS_O_DIRECT;
 649         if (td->o.oatomic) {
 650                 if (!FIO_O_ATOMIC) {
 651                         td_verror(td, EINVAL, "OS does not support atomic IO");
 652                         return 1;
 653                 }
 654                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 655         }
 656         if (td->o.sync_io)
 657                 flags |= O_SYNC;
 658         if (td->o.create_on_open && td->o.allow_create)
 659                 flags |= O_CREAT;
 660 skip_flags:
 661         if (f->filetype != FIO_TYPE_FILE)
 662                 flags |= FIO_O_NOATIME;
 663
 664 open_again:
 665         if (td_write(td)) {
 666                 if (!read_only)
 667                         flags |= O_RDWR;
 668
 669                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 670                         flags |= O_CREAT;
 671
 672                 if (is_std)
 673                         f->fd = dup(STDOUT_FILENO);
 674                 else
 675                         from_hash = file_lookup_open(f, flags);
 676         } else if (td_read(td)) {
 677                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 678                         flags |= O_RDWR;
 679                 else
 680                         flags |= O_RDONLY;
 681
 682                 if (is_std)
 683                         f->fd = dup(STDIN_FILENO);
 684                 else
 685                         from_hash = file_lookup_open(f, flags);
 686         } else if (td_trim(td)) {
 687                 assert(!td_rw(td)); /* should have matched above */
 688                 flags |= O_RDWR;
 689                 from_hash = file_lookup_open(f, flags);
 690         }
 691
 692         if (f->fd == -1) {
 693                 char buf[FIO_VERROR_SIZE];
 694                 int __e = errno;
 695
 696                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 697                         flags &= ~FIO_O_NOATIME;
 698                         goto open_again;
 699                 }
 700                 if (__e == EMFILE && file_close_shadow_fds(td))
 701                         goto open_again;
 702
 703                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 704
 705                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 706                         log_err("fio: looks like your file system does not " \
 707                                 "support direct=1/buffered=0\n");
 708                 }
 709
 710                 td_verror(td, __e, buf);
 711                 return 1;
 712         }
 713
 714         if (!from_hash && f->fd != -1) {
 715                 if (add_file_hash(f)) {
 716                         int fio_unused ret;
 717
 718                         /*
 719                          * Stash away descriptor for later close. This is to
 720                          * work-around a "feature" on Linux, where a close of
 721                          * an fd that has been opened for write will trigger
 722                          * udev to call blkid to check partitions, fs id, etc.
 723                          * That pollutes the device cache, which can slow down
 724                          * unbuffered accesses.
 725                          */
 726                         if (f->shadow_fd == -1)
 727                                 f->shadow_fd = f->fd;
 728                         else {
 729                                 /*
 730                                  * OK to ignore, we haven't done anything
 731                                  * with it
 732                                  */
 733                                 ret = generic_close_file(td, f);
 734                         }
 735                         goto open_again;
 736                 }
 737         }
 738
 739         return 0;
 740 }
 741
 742 /*
 743  * This function i.e. get_file_size() is the default .get_file_size
 744  * implementation of majority of I/O engines.
 745  */
 746 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 747 {
 748         return get_file_size(td, f);
 749 }
 750
 751 /*
 752  * open/close all files, so that ->real_file_size gets set
 753  */
 754 static int get_file_sizes(struct thread_data *td)
 755 {
 756         struct fio_file *f;
 757         unsigned int i;
 758         int err = 0;
 759
 760         for_each_file(td, f, i) {
 761                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 762                                                                 f->file_name);
 763
 764                 if (td_io_get_file_size(td, f)) {
 765                         if (td->error != ENOENT) {
 766                                 log_err("%s\n", td->verror);
 767                                 err = 1;
 768                                 break;
 769                         }
 770                         clear_error(td);
 771                 }
 772
 773                 /*
 774                  * There are corner cases where we end up with -1 for
 775                  * ->real_file_size due to unsupported file type, etc.
 776                  * We then just set to size option value divided by number
 777                  * of files, similar to the way file ->io_size is set.
 778                  * stat(2) failure doesn't set ->real_file_size to -1.
 779                  */
 780                 if (f->real_file_size == -1ULL && td->o.size)
 781                         f->real_file_size = td->o.size / td->o.nr_files;
 782         }
 783
 784         return err;
 785 }
 786
 787 struct fio_mount {
 788         struct flist_head list;
 789         const char *base;
 790         char __base[256];
 791         unsigned int key;
 792 };
 793
 794 /*
 795  * Get free number of bytes for each file on each unique mount.
 796  */
 797 static unsigned long long get_fs_free_counts(struct thread_data *td)
 798 {
 799         struct flist_head *n, *tmp;
 800         unsigned long long ret = 0;
 801         struct fio_mount *fm;
 802         FLIST_HEAD(list);
 803         struct fio_file *f;
 804         unsigned int i;
 805
 806         for_each_file(td, f, i) {
 807                 struct stat sb;
 808                 char buf[256];
 809
 810                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 811                         if (f->real_file_size != -1ULL)
 812                                 ret += f->real_file_size;
 813                         continue;
 814                 } else if (f->filetype != FIO_TYPE_FILE)
 815                         continue;
 816
 817                 buf[255] = '\0';
 818                 strncpy(buf, f->file_name, 255);
 819
 820                 if (stat(buf, &sb) < 0) {
 821                         if (errno != ENOENT)
 822                                 break;
 823                         strcpy(buf, ".");
 824                         if (stat(buf, &sb) < 0)
 825                                 break;
 826                 }
 827
 828                 fm = NULL;
 829                 flist_for_each(n, &list) {
 830                         fm = flist_entry(n, struct fio_mount, list);
 831                         if (fm->key == sb.st_dev)
 832                                 break;
 833
 834                         fm = NULL;
 835                 }
 836
 837                 if (fm)
 838                         continue;
 839
 840                 fm = calloc(1, sizeof(*fm));
 841                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 842                 fm->base = basename(fm->__base);
 843                 fm->key = sb.st_dev;
 844                 flist_add(&fm->list, &list);
 845         }
 846
 847         flist_for_each_safe(n, tmp, &list) {
 848                 unsigned long long sz;
 849
 850                 fm = flist_entry(n, struct fio_mount, list);
 851                 flist_del(&fm->list);
 852
 853                 sz = get_fs_free_size(fm->base);
 854                 if (sz && sz != -1ULL)
 855                         ret += sz;
 856
 857                 free(fm);
 858         }
 859
 860         return ret;
 861 }
 862
 863 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 864 {
 865         struct thread_options *o = &td->o;
 866         unsigned long long align_bs;
 867         unsigned long long offset;
 868
 869         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 870                 return f->real_file_size;
 871
 872         if (o->start_offset_percent > 0) {
 873                 /*
 874                  * if offset_align is provided, set initial offset
 875                  */
 876                 if (fio_option_is_set(o, start_offset_align)) {
 877                         align_bs = o->start_offset_align;
 878                 } else {
 879                         /* else take the minimum block size */
 880                         align_bs = td_min_bs(td);
 881                 }
 882
 883                 /* calculate the raw offset */
 884                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 885                         (td->subjob_number * o->offset_increment);
 886
 887                 /*
 888                  * block align the offset at the next available boundary at
 889                  * ceiling(offset / align_bs) * align_bs
 890                  */
 891                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 892
 893         } else {
 894                 /* start_offset_percent not set */
 895                 offset = o->start_offset +
 896                                 td->subjob_number * o->offset_increment;
 897         }
 898
 899         return offset;
 900 }
 901
 902 /*
 903  * Open the files and setup files sizes, creating files if necessary.
 904  */
 905 int setup_files(struct thread_data *td)
 906 {
 907         unsigned long long total_size, extend_size;
 908         struct thread_options *o = &td->o;
 909         struct fio_file *f;
 910         unsigned int i, nr_fs_extra = 0;
 911         int err = 0, need_extend;
 912         int old_state;
 913         const unsigned int bs = td_min_bs(td);
 914         uint64_t fs = 0;
 915
 916         dprint(FD_FILE, "setup files\n");
 917
 918         old_state = td_bump_runstate(td, TD_SETTING_UP);
 919
 920         if (o->read_iolog_file)
 921                 goto done;
 922
 923         /*
 924          * Find out physical size of files or devices for this thread,
 925          * before we determine I/O size and range of our targets.
 926          * If ioengine defines a setup() method, it's responsible for
 927          * opening the files and setting f->real_file_size to indicate
 928          * the valid range for that file.
 929          */
 930         if (td->io_ops->setup)
 931                 err = td->io_ops->setup(td);
 932         else
 933                 err = get_file_sizes(td);
 934
 935         if (err)
 936                 goto err_out;
 937
 938         /*
 939          * check sizes. if the files/devices do not exist and the size
 940          * isn't passed to fio, abort.
 941          */
 942         total_size = 0;
 943         for_each_file(td, f, i) {
 944                 f->fileno = i;
 945                 if (f->real_file_size == -1ULL)
 946                         total_size = -1ULL;
 947                 else
 948                         total_size += f->real_file_size;
 949         }
 950
 951         if (o->fill_device)
 952                 td->fill_device_size = get_fs_free_counts(td);
 953
 954         /*
 955          * device/file sizes are zero and no size given, punt
 956          */
 957         if ((!total_size || total_size == -1ULL) && !o->size &&
 958             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 959             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 960                 log_err("%s: you need to specify size=\n", o->name);
 961                 td_verror(td, EINVAL, "total_file_size");
 962                 goto err_out;
 963         }
 964
 965         /*
 966          * Calculate per-file size and potential extra size for the
 967          * first files, if needed (i.e. if we don't have a fixed size).
 968          */
 969         if (!o->file_size_low && o->nr_files) {
 970                 uint64_t all_fs;
 971
 972                 fs = o->size / o->nr_files;
 973                 all_fs = fs * o->nr_files;
 974
 975                 if (all_fs < o->size)
 976                         nr_fs_extra = (o->size - all_fs) / bs;
 977         }
 978
 979         /*
 980          * now file sizes are known, so we can set ->io_size. if size= is
 981          * not given, ->io_size is just equal to ->real_file_size. if size
 982          * is given, ->io_size is size / nr_files.
 983          */
 984         extend_size = total_size = 0;
 985         need_extend = 0;
 986         for_each_file(td, f, i) {
 987                 f->file_offset = get_start_offset(td, f);
 988
 989                 /*
 990                  * Update ->io_size depending on options specified.
 991                  * ->file_size_low being 0 means filesize option isn't set.
 992                  * Non zero ->file_size_low equals ->file_size_high means
 993                  * filesize option is set in a fixed size format.
 994                  * Non zero ->file_size_low not equals ->file_size_high means
 995                  * filesize option is set in a range format.
 996                  */
 997                 if (!o->file_size_low) {
 998                         /*
 999                          * no file size or range given, file size is equal to
1000                          * total size divided by number of files. If the size
1001                          * doesn't divide nicely with the min blocksize,
1002                          * make the first files bigger.
1003                          */
1004                         f->io_size = fs;
1005                         if (nr_fs_extra) {
1006                                 nr_fs_extra--;
1007                                 f->io_size += bs;
1008                         }
1009
1010                         /*
1011                          * We normally don't come here for regular files, but
1012                          * if the result is 0 for a regular file, set it to the
1013                          * real file size. This could be size of the existing
1014                          * one if it already exists, but otherwise will be set
1015                          * to 0. A new file won't be created because
1016                          * ->io_size + ->file_offset equals ->real_file_size.
1017                          */
1018                         if (!f->io_size) {
1019                                 if (f->file_offset > f->real_file_size)
1020                                         goto err_offset;
1021                                 f->io_size = f->real_file_size - f->file_offset;
1022                                 if (!f->io_size)
1023                                         log_info("fio: file %s may be ignored\n",
1024                                                 f->file_name);
1025                         }
1026                 } else if (f->real_file_size < o->file_size_low ||
1027                            f->real_file_size > o->file_size_high) {
1028                         if (f->file_offset > o->file_size_low)
1029                                 goto err_offset;
1030                         /*
1031                          * file size given. if it's fixed, use that. if it's a
1032                          * range, generate a random size in-between.
1033                          */
1034                         if (o->file_size_low == o->file_size_high)
1035                                 f->io_size = o->file_size_low - f->file_offset;
1036                         else {
1037                                 f->io_size = get_rand_file_size(td)
1038                                                 - f->file_offset;
1039                         }
1040                 } else
1041                         f->io_size = f->real_file_size - f->file_offset;
1042
1043                 if (f->io_size == -1ULL)
1044                         total_size = -1ULL;
1045                 else {
1046                         if (o->size_percent) {
1047                                 uint64_t file_size;
1048
1049                                 file_size = f->io_size + f->file_offset;
1050                                 f->io_size = (file_size *
1051                                               o->size_percent) / 100;
1052                                 if (f->io_size > (file_size - f->file_offset))
1053                                         f->io_size = file_size - f->file_offset;
1054
1055                                 f->io_size -= (f->io_size % td_min_bs(td));
1056                         }
1057                         total_size += f->io_size;
1058                 }
1059
1060                 if (f->filetype == FIO_TYPE_FILE &&
1061                     (f->io_size + f->file_offset) > f->real_file_size &&
1062                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1063                         if (!o->create_on_open) {
1064                                 need_extend++;
1065                                 extend_size += (f->io_size + f->file_offset);
1066                                 fio_file_set_extend(f);
1067                         } else
1068                                 f->real_file_size = f->io_size + f->file_offset;
1069                 }
1070         }
1071
1072         if (td->o.block_error_hist) {
1073                 int len;
1074
1075                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1076                 f = td->files[0];
1077                 len = f->io_size / td->o.bs[DDIR_TRIM];
1078                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1079                         log_err("fio: cannot calculate block histogram with "
1080                                 "%d trim blocks, maximum %d\n",
1081                                 len, MAX_NR_BLOCK_INFOS);
1082                         td_verror(td, EINVAL, "block_error_hist");
1083                         goto err_out;
1084                 }
1085
1086                 td->ts.nr_block_infos = len;
1087                 for (i = 0; i < len; i++)
1088                         td->ts.block_infos[i] =
1089                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1090         } else
1091                 td->ts.nr_block_infos = 0;
1092
1093         if (!o->size || (total_size && o->size > total_size))
1094                 o->size = total_size;
1095
1096         if (o->size < td_min_bs(td)) {
1097                 log_err("fio: blocksize too large for data set\n");
1098                 goto err_out;
1099         }
1100
1101         /*
1102          * See if we need to extend some files, typically needed when our
1103          * target regular files don't exist yet, but our jobs require them
1104          * initially due to read I/Os.
1105          */
1106         if (need_extend) {
1107                 temp_stall_ts = 1;
1108                 if (output_format & FIO_OUTPUT_NORMAL) {
1109                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1110                                  o->name,
1111                                  need_extend > 1 ? "s" : "",
1112                                  need_extend,
1113                                  need_extend > 1 ? "s" : "",
1114                                  need_extend > 1 ? "total " : "",
1115                                  extend_size >> 20);
1116                 }
1117
1118                 for_each_file(td, f, i) {
1119                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1120
1121                         if (!fio_file_extend(f))
1122                                 continue;
1123
1124                         assert(f->filetype == FIO_TYPE_FILE);
1125                         fio_file_clear_extend(f);
1126                         if (!o->fill_device) {
1127                                 old_len = f->real_file_size;
1128                                 extend_len = f->io_size + f->file_offset -
1129                                                 old_len;
1130                         }
1131                         f->real_file_size = (f->io_size + f->file_offset);
1132                         err = extend_file(td, f);
1133                         if (err)
1134                                 break;
1135
1136                         err = __file_invalidate_cache(td, f, old_len,
1137                                                                 extend_len);
1138
1139                         /*
1140                          * Shut up static checker
1141                          */
1142                         if (f->fd != -1)
1143                                 close(f->fd);
1144
1145                         f->fd = -1;
1146                         if (err)
1147                                 break;
1148                 }
1149                 temp_stall_ts = 0;
1150         }
1151
1152         if (err)
1153                 goto err_out;
1154
1155         if (!o->zone_size)
1156                 o->zone_size = o->size;
1157
1158         /*
1159          * iolog already set the total io size, if we read back
1160          * stored entries.
1161          */
1162         if (!o->read_iolog_file) {
1163                 if (o->io_size)
1164                         td->total_io_size = o->io_size * o->loops;
1165                 else
1166                         td->total_io_size = o->size * o->loops;
1167         }
1168
1169 done:
1170         if (o->create_only)
1171                 td->done = 1;
1172
1173         td_restore_runstate(td, old_state);
1174         return 0;
1175 err_offset:
1176         log_err("%s: you need to specify valid offset=\n", o->name);
1177 err_out:
1178         td_restore_runstate(td, old_state);
1179         return 1;
1180 }
1181
1182 bool pre_read_files(struct thread_data *td)
1183 {
1184         struct fio_file *f;
1185         unsigned int i;
1186
1187         dprint(FD_FILE, "pre_read files\n");
1188
1189         for_each_file(td, f, i) {
1190                 if (!pre_read_file(td, f))
1191                         return false;
1192         }
1193
1194         return true;
1195 }
1196
1197 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1198 {
1199         unsigned int range_size, seed;
1200         unsigned long nranges;
1201         uint64_t fsize;
1202
1203         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1204         fsize = min(f->real_file_size, f->io_size);
1205
1206         nranges = (fsize + range_size - 1) / range_size;
1207
1208         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1209         if (!td->o.rand_repeatable)
1210                 seed = td->rand_seeds[4];
1211
1212         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1213                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1214         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1215                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1216         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1217                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1218 }
1219
1220 static bool 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 false;
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         return true;
1236 }
1237
1238 /*
1239  * Check if the number of blocks exceeds the randomness capability of
1240  * the selected generator. Tausworthe is 32-bit, the others are fullly
1241  * 64-bit capable.
1242  */
1243 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1244                                  uint64_t blocks)
1245 {
1246         if (blocks <= FRAND32_MAX)
1247                 return 0;
1248         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1249                 return 0;
1250
1251         /*
1252          * If the user hasn't specified a random generator, switch
1253          * to tausworthe64 with informational warning. If the user did
1254          * specify one, just warn.
1255          */
1256         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1257                         f->file_name);
1258
1259         if (!fio_option_is_set(&td->o, random_generator)) {
1260                 log_info("fio: Switching to tausworthe64. Use the "
1261                          "random_generator= option to get rid of this "
1262                          "warning.\n");
1263                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1264                 return 0;
1265         }
1266
1267         /*
1268          * Just make this information to avoid breaking scripts.
1269          */
1270         log_info("fio: Use the random_generator= option to switch to lfsr or "
1271                          "tausworthe64.\n");
1272         return 0;
1273 }
1274
1275 bool init_random_map(struct thread_data *td)
1276 {
1277         unsigned long long blocks;
1278         struct fio_file *f;
1279         unsigned int i;
1280
1281         if (init_rand_distribution(td))
1282                 return true;
1283         if (!td_random(td))
1284                 return true;
1285
1286         for_each_file(td, f, i) {
1287                 uint64_t fsize = min(f->real_file_size, f->io_size);
1288
1289                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1290
1291                 if (check_rand_gen_limits(td, f, blocks))
1292                         return false;
1293
1294                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1295                         unsigned long seed;
1296
1297                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1298
1299                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1300                                 fio_file_set_lfsr(f);
1301                                 continue;
1302                         }
1303                 } else if (!td->o.norandommap) {
1304                         f->io_axmap = axmap_new(blocks);
1305                         if (f->io_axmap) {
1306                                 fio_file_set_axmap(f);
1307                                 continue;
1308                         }
1309                 } else if (td->o.norandommap)
1310                         continue;
1311
1312                 if (!td->o.softrandommap) {
1313                         log_err("fio: failed allocating random map. If running"
1314                                 " a large number of jobs, try the 'norandommap'"
1315                                 " option or set 'softrandommap'. Or give"
1316                                 " a larger --alloc-size to fio.\n");
1317                         return false;
1318                 }
1319
1320                 log_info("fio: file %s failed allocating random map. Running "
1321                          "job without.\n", f->file_name);
1322         }
1323
1324         return true;
1325 }
1326
1327 void close_files(struct thread_data *td)
1328 {
1329         struct fio_file *f;
1330         unsigned int i;
1331
1332         for_each_file(td, f, i) {
1333                 if (fio_file_open(f))
1334                         td_io_close_file(td, f);
1335         }
1336 }
1337
1338 void close_and_free_files(struct thread_data *td)
1339 {
1340         struct fio_file *f;
1341         unsigned int i;
1342         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1343
1344         dprint(FD_FILE, "close files\n");
1345
1346         for_each_file(td, f, i) {
1347                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1348                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1349                         td_io_unlink_file(td, f);
1350                 }
1351
1352                 if (fio_file_open(f))
1353                         td_io_close_file(td, f);
1354
1355                 remove_file_hash(f);
1356
1357                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1358                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1359                         td_io_unlink_file(td, f);
1360                 }
1361
1362                 if (use_free)
1363                         free(f->file_name);
1364                 else
1365                         sfree(f->file_name);
1366                 f->file_name = NULL;
1367                 if (fio_file_axmap(f)) {
1368                         axmap_free(f->io_axmap);
1369                         f->io_axmap = NULL;
1370                 }
1371                 if (use_free)
1372                         free(f);
1373                 else
1374                         sfree(f);
1375         }
1376
1377         td->o.filename = NULL;
1378         free(td->files);
1379         free(td->file_locks);
1380         td->files_index = 0;
1381         td->files = NULL;
1382         td->file_locks = NULL;
1383         td->o.file_lock_mode = FILE_LOCK_NONE;
1384         td->o.nr_files = 0;
1385 }
1386
1387 static void get_file_type(struct fio_file *f)
1388 {
1389         struct stat sb;
1390
1391         if (!strcmp(f->file_name, "-"))
1392                 f->filetype = FIO_TYPE_PIPE;
1393         else
1394                 f->filetype = FIO_TYPE_FILE;
1395
1396 #ifdef WIN32
1397         /* \\.\ is the device namespace in Windows, where every file is
1398          * a block device */
1399         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1400                 f->filetype = FIO_TYPE_BLOCK;
1401 #endif
1402
1403         if (!stat(f->file_name, &sb)) {
1404                 if (S_ISBLK(sb.st_mode))
1405                         f->filetype = FIO_TYPE_BLOCK;
1406                 else if (S_ISCHR(sb.st_mode))
1407                         f->filetype = FIO_TYPE_CHAR;
1408                 else if (S_ISFIFO(sb.st_mode))
1409                         f->filetype = FIO_TYPE_PIPE;
1410         }
1411 }
1412
1413 static bool __is_already_allocated(const char *fname, bool set)
1414 {
1415         struct flist_head *entry;
1416         bool ret;
1417
1418         ret = file_bloom_exists(fname, set);
1419         if (!ret)
1420                 return ret;
1421
1422         flist_for_each(entry, &filename_list) {
1423                 struct file_name *fn;
1424
1425                 fn = flist_entry(entry, struct file_name, list);
1426
1427                 if (!strcmp(fn->filename, fname))
1428                         return true;
1429         }
1430
1431         return false;
1432 }
1433
1434 static bool is_already_allocated(const char *fname)
1435 {
1436         bool ret;
1437
1438         fio_file_hash_lock();
1439         ret = __is_already_allocated(fname, false);
1440         fio_file_hash_unlock();
1441
1442         return ret;
1443 }
1444
1445 static void set_already_allocated(const char *fname)
1446 {
1447         struct file_name *fn;
1448
1449         fn = malloc(sizeof(struct file_name));
1450         fn->filename = strdup(fname);
1451
1452         fio_file_hash_lock();
1453         if (!__is_already_allocated(fname, true)) {
1454                 flist_add_tail(&fn->list, &filename_list);
1455                 fn = NULL;
1456         }
1457         fio_file_hash_unlock();
1458
1459         if (fn) {
1460                 free(fn->filename);
1461                 free(fn);
1462         }
1463 }
1464
1465 static void free_already_allocated(void)
1466 {
1467         struct flist_head *entry, *tmp;
1468         struct file_name *fn;
1469
1470         if (flist_empty(&filename_list))
1471                 return;
1472
1473         fio_file_hash_lock();
1474         flist_for_each_safe(entry, tmp, &filename_list) {
1475                 fn = flist_entry(entry, struct file_name, list);
1476                 free(fn->filename);
1477                 flist_del(&fn->list);
1478                 free(fn);
1479         }
1480
1481         fio_file_hash_unlock();
1482 }
1483
1484 static struct fio_file *alloc_new_file(struct thread_data *td)
1485 {
1486         struct fio_file *f;
1487
1488         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1489                 f = calloc(1, sizeof(*f));
1490         else
1491                 f = smalloc(sizeof(*f));
1492         if (!f) {
1493                 assert(0);
1494                 return NULL;
1495         }
1496
1497         f->fd = -1;
1498         f->shadow_fd = -1;
1499         fio_file_reset(td, f);
1500         return f;
1501 }
1502
1503 bool exists_and_not_regfile(const char *filename)
1504 {
1505         struct stat sb;
1506
1507         if (lstat(filename, &sb) == -1)
1508                 return false;
1509
1510 #ifndef WIN32 /* NOT Windows */
1511         if (S_ISREG(sb.st_mode))
1512                 return false;
1513 #else
1514         /* \\.\ is the device namespace in Windows, where every file
1515          * is a device node */
1516         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1517                 return false;
1518 #endif
1519
1520         return true;
1521 }
1522
1523 static bool create_work_dirs(struct thread_data *td, const char *fname)
1524 {
1525         char path[PATH_MAX];
1526         char *start, *end;
1527
1528         if (td->o.directory) {
1529                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1530                          FIO_OS_PATH_SEPARATOR, fname);
1531                 start = strstr(path, fname);
1532         } else {
1533                 snprintf(path, PATH_MAX, "%s", fname);
1534                 start = path;
1535         }
1536
1537         end = start;
1538         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1539                 if (end == start)
1540                         break;
1541                 *end = '\0';
1542                 errno = 0;
1543 #ifdef CONFIG_HAVE_MKDIR_TWO
1544                 if (mkdir(path, 0600) && errno != EEXIST) {
1545 #else
1546                 if (mkdir(path) && errno != EEXIST) {
1547 #endif
1548                         log_err("fio: failed to create dir (%s): %d\n",
1549                                 start, errno);
1550                         return false;
1551                 }
1552                 *end = FIO_OS_PATH_SEPARATOR;
1553                 end++;
1554         }
1555         td->flags |= TD_F_DIRS_CREATED;
1556         return true;
1557 }
1558
1559 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1560 {
1561         int cur_files = td->files_index;
1562         char file_name[PATH_MAX];
1563         struct fio_file *f;
1564         int len = 0;
1565
1566         dprint(FD_FILE, "add file %s\n", fname);
1567
1568         if (td->o.directory)
1569                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1570                                         td->o.unique_filename);
1571
1572         sprintf(file_name + len, "%s", fname);
1573
1574         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1575             !(td->flags & TD_F_DIRS_CREATED) &&
1576             !create_work_dirs(td, fname))
1577                 return 1;
1578
1579         /* clean cloned siblings using existing files */
1580         if (numjob && is_already_allocated(file_name) &&
1581             !exists_and_not_regfile(fname))
1582                 return 0;
1583
1584         f = alloc_new_file(td);
1585
1586         if (td->files_size <= td->files_index) {
1587                 unsigned int new_size = td->o.nr_files + 1;
1588
1589                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1590
1591                 td->files = realloc(td->files, new_size * sizeof(f));
1592                 if (td->files == NULL) {
1593                         log_err("fio: realloc OOM\n");
1594                         assert(0);
1595                 }
1596                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1597                         td->file_locks = realloc(td->file_locks, new_size);
1598                         if (!td->file_locks) {
1599                                 log_err("fio: realloc OOM\n");
1600                                 assert(0);
1601                         }
1602                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1603                 }
1604                 td->files_size = new_size;
1605         }
1606         td->files[cur_files] = f;
1607         f->fileno = cur_files;
1608
1609         /*
1610          * init function, io engine may not be loaded yet
1611          */
1612         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1613                 f->real_file_size = -1ULL;
1614
1615         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1616                 f->file_name = strdup(file_name);
1617         else
1618                 f->file_name = smalloc_strdup(file_name);
1619         if (!f->file_name)
1620                 assert(0);
1621
1622         get_file_type(f);
1623
1624         switch (td->o.file_lock_mode) {
1625         case FILE_LOCK_NONE:
1626                 break;
1627         case FILE_LOCK_READWRITE:
1628                 f->rwlock = fio_rwlock_init();
1629                 break;
1630         case FILE_LOCK_EXCLUSIVE:
1631                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1632                 break;
1633         default:
1634                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1635                 assert(0);
1636         }
1637
1638         td->files_index++;
1639         if (f->filetype == FIO_TYPE_FILE)
1640                 td->nr_normal_files++;
1641
1642         if (td->o.numjobs > 1)
1643                 set_already_allocated(file_name);
1644
1645         if (inc)
1646                 td->o.nr_files++;
1647
1648         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1649                                                         cur_files);
1650
1651         return cur_files;
1652 }
1653
1654 int add_file_exclusive(struct thread_data *td, const char *fname)
1655 {
1656         struct fio_file *f;
1657         unsigned int i;
1658
1659         for_each_file(td, f, i) {
1660                 if (!strcmp(f->file_name, fname))
1661                         return i;
1662         }
1663
1664         return add_file(td, fname, 0, 1);
1665 }
1666
1667 void get_file(struct fio_file *f)
1668 {
1669         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1670         assert(fio_file_open(f));
1671         f->references++;
1672 }
1673
1674 int put_file(struct thread_data *td, struct fio_file *f)
1675 {
1676         int f_ret = 0, ret = 0;
1677
1678         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1679
1680         if (!fio_file_open(f)) {
1681                 assert(f->fd == -1);
1682                 return 0;
1683         }
1684
1685         assert(f->references);
1686         if (--f->references)
1687                 return 0;
1688
1689         if (should_fsync(td) && td->o.fsync_on_close) {
1690                 f_ret = fsync(f->fd);
1691                 if (f_ret < 0)
1692                         f_ret = errno;
1693         }
1694
1695         if (td->io_ops->close_file)
1696                 ret = td->io_ops->close_file(td, f);
1697
1698         if (!ret)
1699                 ret = f_ret;
1700
1701         td->nr_open_files--;
1702         fio_file_clear_open(f);
1703         assert(f->fd == -1);
1704         return ret;
1705 }
1706
1707 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1708 {
1709         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1710                 return;
1711
1712         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1713                 if (ddir == DDIR_READ)
1714                         fio_rwlock_read(f->rwlock);
1715                 else
1716                         fio_rwlock_write(f->rwlock);
1717         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1718                 fio_mutex_down(f->lock);
1719
1720         td->file_locks[f->fileno] = td->o.file_lock_mode;
1721 }
1722
1723 void unlock_file(struct thread_data *td, struct fio_file *f)
1724 {
1725         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1726                 return;
1727
1728         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1729                 fio_rwlock_unlock(f->rwlock);
1730         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1731                 fio_mutex_up(f->lock);
1732
1733         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1734 }
1735
1736 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1737 {
1738         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1739                 return;
1740         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1741                 unlock_file(td, f);
1742 }
1743
1744 static bool recurse_dir(struct thread_data *td, const char *dirname)
1745 {
1746         struct dirent *dir;
1747         bool ret = false;
1748         DIR *D;
1749
1750         D = opendir(dirname);
1751         if (!D) {
1752                 char buf[FIO_VERROR_SIZE];
1753
1754                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1755                 td_verror(td, errno, buf);
1756                 return true;
1757         }
1758
1759         while ((dir = readdir(D)) != NULL) {
1760                 char full_path[PATH_MAX];
1761                 struct stat sb;
1762
1763                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1764                         continue;
1765
1766                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1767
1768                 if (lstat(full_path, &sb) == -1) {
1769                         if (errno != ENOENT) {
1770                                 td_verror(td, errno, "stat");
1771                                 ret = true;
1772                                 break;
1773                         }
1774                 }
1775
1776                 if (S_ISREG(sb.st_mode)) {
1777                         add_file(td, full_path, 0, 1);
1778                         continue;
1779                 }
1780                 if (!S_ISDIR(sb.st_mode))
1781                         continue;
1782
1783                 ret = recurse_dir(td, full_path);
1784                 if (ret)
1785                         break;
1786         }
1787
1788         closedir(D);
1789         return ret;
1790 }
1791
1792 int add_dir_files(struct thread_data *td, const char *path)
1793 {
1794         int ret = recurse_dir(td, path);
1795
1796         if (!ret)
1797                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1798
1799         return ret;
1800 }
1801
1802 void dup_files(struct thread_data *td, struct thread_data *org)
1803 {
1804         struct fio_file *f;
1805         unsigned int i;
1806
1807         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1808
1809         if (!org->files)
1810                 return;
1811
1812         td->files = malloc(org->files_index * sizeof(f));
1813
1814         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1815                 td->file_locks = malloc(org->files_index);
1816
1817         for_each_file(org, f, i) {
1818                 struct fio_file *__f;
1819
1820                 __f = alloc_new_file(td);
1821
1822                 if (f->file_name) {
1823                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1824                                 __f->file_name = strdup(f->file_name);
1825                         else
1826                                 __f->file_name = smalloc_strdup(f->file_name);
1827                         if (!__f->file_name)
1828                                 assert(0);
1829
1830                         __f->filetype = f->filetype;
1831                 }
1832
1833                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1834                         __f->lock = f->lock;
1835                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1836                         __f->rwlock = f->rwlock;
1837
1838                 td->files[i] = __f;
1839         }
1840 }
1841
1842 /*
1843  * Returns the index that matches the filename, or -1 if not there
1844  */
1845 int get_fileno(struct thread_data *td, const char *fname)
1846 {
1847         struct fio_file *f;
1848         unsigned int i;
1849
1850         for_each_file(td, f, i)
1851                 if (!strcmp(f->file_name, fname))
1852                         return i;
1853
1854         return -1;
1855 }
1856
1857 /*
1858  * For log usage, where we add/open/close files automatically
1859  */
1860 void free_release_files(struct thread_data *td)
1861 {
1862         close_files(td);
1863         td->o.nr_files = 0;
1864         td->o.open_files = 0;
1865         td->files_index = 0;
1866         td->nr_normal_files = 0;
1867 }
1868
1869 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1870 {
1871         int i;
1872
1873         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1874                 f->last_pos[i] = f->file_offset;
1875                 f->last_start[i] = -1ULL;
1876         }
1877
1878         if (fio_file_axmap(f))
1879                 axmap_reset(f->io_axmap);
1880         else if (fio_file_lfsr(f))
1881                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1882 }
1883
1884 bool fio_files_done(struct thread_data *td)
1885 {
1886         struct fio_file *f;
1887         unsigned int i;
1888
1889         for_each_file(td, f, i)
1890                 if (!fio_file_done(f))
1891                         return false;
1892
1893         return true;
1894 }
1895
1896 /* free memory used in initialization phase only */
1897 void filesetup_mem_free(void)
1898 {
1899         free_already_allocated();
1900 }
1901
1902 /*
1903  * This function is for platforms which support direct I/O but not O_DIRECT.
1904  */
1905 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1906 {
1907 #ifdef FIO_OS_DIRECTIO
1908         int ret = fio_set_odirect(f);
1909
1910         if (ret) {
1911                 td_verror(td, ret, "fio_set_directio");
1912 #if defined(__sun__)
1913                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1914                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1915                 } else {
1916                         log_err("fio: the file system does not seem to support direct IO\n");
1917                 }
1918 #else
1919                 log_err("fio: the file system does not seem to support direct IO\n");
1920 #endif
1921                 return -1;
1922         }
1923
1924         return 0;
1925 #else
1926         log_err("fio: direct IO is not supported on this host operating system\n");
1927         return -1;
1928 #endif
1929 }