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
   9 #include "fio.h"
  10 #include "smalloc.h"
  11 #include "filehash.h"
  12 #include "options.h"
  13 #include "os/os.h"
  14 #include "hash.h"
  15 #include "lib/axmap.h"
  16 #include "rwlock.h"
  17
  18 #ifdef CONFIG_LINUX_FALLOCATE
  19 #include <linux/falloc.h>
  20 #endif
  21
  22 static FLIST_HEAD(filename_list);
  23
  24 /*
  25  * List entry for filename_list
  26  */
  27 struct file_name {
  28         struct flist_head list;
  29         char *filename;
  30 };
  31
  32 static inline void clear_error(struct thread_data *td)
  33 {
  34         td->error = 0;
  35         td->verror[0] = '\0';
  36 }
  37
  38 static int native_fallocate(struct thread_data *td, struct fio_file *f)
  39 {
  40         bool success;
  41
  42         success = fio_fallocate(f, 0, f->real_file_size);
  43         dprint(FD_FILE, "native fallocate of file %s size %llu was "
  44                         "%ssuccessful\n", f->file_name,
  45                         (unsigned long long) f->real_file_size,
  46                         !success ? "un": "");
  47
  48         if (success)
  49                 return false;
  50
  51         if (errno == ENOSYS)
  52                 dprint(FD_FILE, "native fallocate is not implemented\n");
  53
  54         return true;
  55 }
  56
  57 static void fallocate_file(struct thread_data *td, struct fio_file *f)
  58 {
  59         if (td->o.fill_device)
  60                 return;
  61
  62         switch (td->o.fallocate_mode) {
  63         case FIO_FALLOCATE_NATIVE:
  64                 native_fallocate(td, f);
  65                 break;
  66         case FIO_FALLOCATE_NONE:
  67                 break;
  68 #ifdef CONFIG_POSIX_FALLOCATE
  69         case FIO_FALLOCATE_POSIX: {
  70                 int r;
  71
  72                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  73                                  f->file_name,
  74                                  (unsigned long long) f->real_file_size);
  75
  76                 r = posix_fallocate(f->fd, 0, f->real_file_size);
  77                 if (r > 0)
  78                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
  79                 break;
  80                 }
  81 #endif /* CONFIG_POSIX_FALLOCATE */
  82 #ifdef CONFIG_LINUX_FALLOCATE
  83         case FIO_FALLOCATE_KEEP_SIZE: {
  84                 int r;
  85
  86                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
  87                                 "file %s size %llu\n", f->file_name,
  88                                 (unsigned long long) f->real_file_size);
  89
  90                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
  91                 if (r != 0)
  92                         td_verror(td, errno, "fallocate");
  93
  94                 break;
  95                 }
  96 #endif /* CONFIG_LINUX_FALLOCATE */
  97         default:
  98                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
  99                 assert(0);
 100         }
 101 }
 102
 103 /*
 104  * Leaves f->fd open on success, caller must close
 105  */
 106 static int extend_file(struct thread_data *td, struct fio_file *f)
 107 {
 108         int new_layout = 0, unlink_file = 0, flags;
 109         unsigned long long left;
 110         unsigned int bs;
 111         char *b = NULL;
 112
 113         if (read_only) {
 114                 log_err("fio: refusing extend of file due to read-only\n");
 115                 return 0;
 116         }
 117
 118         /*
 119          * check if we need to lay the file out complete again. fio
 120          * does that for operations involving reads, or for writes
 121          * where overwrite is set
 122          */
 123         if (td_read(td) ||
 124            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
 125             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
 126                 new_layout = 1;
 127         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
 128                 unlink_file = 1;
 129
 130         if (unlink_file || new_layout) {
 131                 int ret;
 132
 133                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
 134
 135                 ret = td_io_unlink_file(td, f);
 136                 if (ret != 0 && ret != ENOENT) {
 137                         td_verror(td, errno, "unlink");
 138                         return 1;
 139                 }
 140         }
 141
 142         flags = O_WRONLY;
 143         if (td->o.allow_create)
 144                 flags |= O_CREAT;
 145         if (new_layout)
 146                 flags |= O_TRUNC;
 147
 148 #ifdef WIN32
 149         flags |= _O_BINARY;
 150 #endif
 151
 152         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 153         f->fd = open(f->file_name, flags, 0644);
 154         if (f->fd < 0) {
 155                 int err = errno;
 156
 157                 if (err == ENOENT && !td->o.allow_create)
 158                         log_err("fio: file creation disallowed by "
 159                                         "allow_file_create=0\n");
 160                 else
 161                         td_verror(td, err, "open");
 162                 return 1;
 163         }
 164
 165         fallocate_file(td, f);
 166
 167         /*
 168          * If our jobs don't require regular files initially, we're done.
 169          */
 170         if (!new_layout)
 171                 goto done;
 172
 173         /*
 174          * The size will be -1ULL when fill_device is used, so don't truncate
 175          * or fallocate this file, just write it
 176          */
 177         if (!td->o.fill_device) {
 178                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 179                                         (unsigned long long) f->real_file_size);
 180                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 181                         if (errno != EFBIG) {
 182                                 td_verror(td, errno, "ftruncate");
 183                                 goto err;
 184                         }
 185                 }
 186         }
 187
 188         left = f->real_file_size;
 189         bs = td->o.max_bs[DDIR_WRITE];
 190         if (bs > left)
 191                 bs = left;
 192
 193         b = malloc(bs);
 194         if (!b) {
 195                 td_verror(td, errno, "malloc");
 196                 goto err;
 197         }
 198
 199         while (left && !td->terminate) {
 200                 ssize_t r;
 201
 202                 if (bs > left)
 203                         bs = left;
 204
 205                 fill_io_buffer(td, b, bs, bs);
 206
 207                 r = write(f->fd, b, bs);
 208
 209                 if (r > 0) {
 210                         left -= r;
 211                         continue;
 212                 } else {
 213                         if (r < 0) {
 214                                 int __e = errno;
 215
 216                                 if (__e == ENOSPC) {
 217                                         if (td->o.fill_device)
 218                                                 break;
 219                                         log_info("fio: ENOSPC on laying out "
 220                                                  "file, stopping\n");
 221                                         break;
 222                                 }
 223                                 td_verror(td, errno, "write");
 224                         } else
 225                                 td_verror(td, EIO, "write");
 226
 227                         break;
 228                 }
 229         }
 230
 231         if (td->terminate) {
 232                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 233                 td_io_unlink_file(td, f);
 234         } else if (td->o.create_fsync) {
 235                 if (fsync(f->fd) < 0) {
 236                         td_verror(td, errno, "fsync");
 237                         goto err;
 238                 }
 239         }
 240         if (td->o.fill_device && !td_write(td)) {
 241                 fio_file_clear_size_known(f);
 242                 if (td_io_get_file_size(td, f))
 243                         goto err;
 244                 if (f->io_size > f->real_file_size)
 245                         f->io_size = f->real_file_size;
 246         }
 247
 248         free(b);
 249 done:
 250         return 0;
 251 err:
 252         close(f->fd);
 253         f->fd = -1;
 254         if (b)
 255                 free(b);
 256         return 1;
 257 }
 258
 259 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 260 {
 261         int r, did_open = 0, old_runstate;
 262         unsigned long long left;
 263         unsigned int bs;
 264         bool ret = true;
 265         char *b;
 266
 267         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 268             td_ioengine_flagged(td, FIO_NOIO))
 269                 return true;
 270
 271         if (f->filetype == FIO_TYPE_CHAR)
 272                 return true;
 273
 274         if (!fio_file_open(f)) {
 275                 if (td->io_ops->open_file(td, f)) {
 276                         log_err("fio: cannot pre-read, failed to open file\n");
 277                         return false;
 278                 }
 279                 did_open = 1;
 280         }
 281
 282         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 283
 284         left = f->io_size;
 285         bs = td->o.max_bs[DDIR_READ];
 286         if (bs > left)
 287                 bs = left;
 288
 289         b = malloc(bs);
 290         if (!b) {
 291                 td_verror(td, errno, "malloc");
 292                 ret = false;
 293                 goto error;
 294         }
 295         memset(b, 0, bs);
 296
 297         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 298                 td_verror(td, errno, "lseek");
 299                 log_err("fio: failed to lseek pre-read file\n");
 300                 ret = false;
 301                 goto error;
 302         }
 303
 304         while (left && !td->terminate) {
 305                 if (bs > left)
 306                         bs = left;
 307
 308                 r = read(f->fd, b, bs);
 309
 310                 if (r == (int) bs) {
 311                         left -= bs;
 312                         continue;
 313                 } else {
 314                         td_verror(td, EIO, "pre_read");
 315                         break;
 316                 }
 317         }
 318
 319 error:
 320         td_restore_runstate(td, old_runstate);
 321
 322         if (did_open)
 323                 td->io_ops->close_file(td, f);
 324
 325         free(b);
 326         return ret;
 327 }
 328
 329 unsigned long long get_rand_file_size(struct thread_data *td)
 330 {
 331         unsigned long long ret, sized;
 332         uint64_t frand_max;
 333         unsigned long r;
 334
 335         frand_max = rand_max(&td->file_size_state);
 336         r = __rand(&td->file_size_state);
 337         sized = td->o.file_size_high - td->o.file_size_low;
 338         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 339         ret += td->o.file_size_low;
 340         ret -= (ret % td->o.rw_min_bs);
 341         return ret;
 342 }
 343
 344 static int file_size(struct thread_data *td, struct fio_file *f)
 345 {
 346         struct stat st;
 347
 348         if (stat(f->file_name, &st) == -1) {
 349                 td_verror(td, errno, "fstat");
 350                 return 1;
 351         }
 352
 353         f->real_file_size = st.st_size;
 354         return 0;
 355 }
 356
 357 static int bdev_size(struct thread_data *td, struct fio_file *f)
 358 {
 359         unsigned long long bytes = 0;
 360         int r;
 361
 362         if (td->io_ops->open_file(td, f)) {
 363                 log_err("fio: failed opening blockdev %s for size check\n",
 364                         f->file_name);
 365                 return 1;
 366         }
 367
 368         r = blockdev_size(f, &bytes);
 369         if (r) {
 370                 td_verror(td, r, "blockdev_size");
 371                 goto err;
 372         }
 373
 374         if (!bytes) {
 375                 log_err("%s: zero sized block device?\n", f->file_name);
 376                 goto err;
 377         }
 378
 379         f->real_file_size = bytes;
 380         td->io_ops->close_file(td, f);
 381         return 0;
 382 err:
 383         td->io_ops->close_file(td, f);
 384         return 1;
 385 }
 386
 387 static int char_size(struct thread_data *td, struct fio_file *f)
 388 {
 389 #ifdef FIO_HAVE_CHARDEV_SIZE
 390         unsigned long long bytes = 0;
 391         int r;
 392
 393         if (td->io_ops->open_file(td, f)) {
 394                 log_err("fio: failed opening chardev %s for size check\n",
 395                         f->file_name);
 396                 return 1;
 397         }
 398
 399         r = chardev_size(f, &bytes);
 400         if (r) {
 401                 td_verror(td, r, "chardev_size");
 402                 goto err;
 403         }
 404
 405         if (!bytes) {
 406                 log_err("%s: zero sized char device?\n", f->file_name);
 407                 goto err;
 408         }
 409
 410         f->real_file_size = bytes;
 411         td->io_ops->close_file(td, f);
 412         return 0;
 413 err:
 414         td->io_ops->close_file(td, f);
 415         return 1;
 416 #else
 417         f->real_file_size = -1ULL;
 418         return 0;
 419 #endif
 420 }
 421
 422 static int get_file_size(struct thread_data *td, struct fio_file *f)
 423 {
 424         int ret = 0;
 425
 426         if (fio_file_size_known(f))
 427                 return 0;
 428
 429         if (f->filetype == FIO_TYPE_FILE)
 430                 ret = file_size(td, f);
 431         else if (f->filetype == FIO_TYPE_BLOCK)
 432                 ret = bdev_size(td, f);
 433         else if (f->filetype == FIO_TYPE_CHAR)
 434                 ret = char_size(td, f);
 435         else {
 436                 f->real_file_size = -1;
 437                 log_info("%s: failed to get file size of %s\n", td->o.name,
 438                                         f->file_name);
 439                 return 1; /* avoid offset extends end error message */
 440         }
 441
 442         /*
 443          * Leave ->real_file_size with 0 since it could be expectation
 444          * of initial setup for regular files.
 445          */
 446         if (ret)
 447                 return ret;
 448
 449         /*
 450          * ->file_offset normally hasn't been initialized yet, so this
 451          * is basically always false unless ->real_file_size is -1, but
 452          * if ->real_file_size is -1 this message doesn't make sense.
 453          * As a result, this message is basically useless.
 454          */
 455         if (f->file_offset > f->real_file_size) {
 456                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 457                                         (unsigned long long) f->file_offset,
 458                                         (unsigned long long) f->real_file_size);
 459                 return 1;
 460         }
 461
 462         fio_file_set_size_known(f);
 463         return 0;
 464 }
 465
 466 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 467                                    unsigned long long off,
 468                                    unsigned long long len)
 469 {
 470         int errval = 0, ret = 0;
 471
 472 #ifdef CONFIG_ESX
 473         return 0;
 474 #endif
 475
 476         if (len == -1ULL)
 477                 len = f->io_size;
 478         if (off == -1ULL)
 479                 off = f->file_offset;
 480
 481         if (len == -1ULL || off == -1ULL)
 482                 return 0;
 483
 484         if (td->io_ops->invalidate) {
 485                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 486                         f->file_name);
 487                 ret = td->io_ops->invalidate(td, f);
 488                 if (ret < 0)
 489                         errval = -ret;
 490         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 491                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
 492                        td->io_ops->name);
 493                 ret = 0;
 494         } else if (f->filetype == FIO_TYPE_FILE) {
 495                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 496                         f->file_name, off, len);
 497                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 498                 if (ret)
 499                         errval = ret;
 500         } else if (f->filetype == FIO_TYPE_BLOCK) {
 501                 int retry_count = 0;
 502
 503                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 504                 ret = blockdev_invalidate_cache(f);
 505                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 506                         /*
 507                          * Linux multipath devices reject ioctl while
 508                          * the maps are being updated. That window can
 509                          * last tens of milliseconds; we'll try up to
 510                          * a quarter of a second.
 511                          */
 512                         usleep(10000);
 513                         ret = blockdev_invalidate_cache(f);
 514                 }
 515                 if (ret < 0 && errno == EACCES && geteuid()) {
 516                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 517                                 log_err("fio: only root may flush block "
 518                                         "devices. Cache flush bypassed!\n");
 519                         }
 520                         ret = 0;
 521                 }
 522                 if (ret < 0)
 523                         errval = errno;
 524         } else if (f->filetype == FIO_TYPE_CHAR ||
 525                    f->filetype == FIO_TYPE_PIPE) {
 526                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 527                 ret = 0;
 528         }
 529
 530         /*
 531          * Cache flushing isn't a fatal condition, and we know it will
 532          * happen on some platforms where we don't have the proper
 533          * function to flush eg block device caches. So just warn and
 534          * continue on our way.
 535          */
 536         if (errval)
 537                 log_info("fio: cache invalidation of %s failed: %s\n",
 538                          f->file_name, strerror(errval));
 539
 540         return 0;
 541
 542 }
 543
 544 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 545 {
 546         if (!fio_file_open(f))
 547                 return 0;
 548
 549         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 550 }
 551
 552 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 553 {
 554         int ret = 0;
 555
 556         dprint(FD_FILE, "fd close %s\n", f->file_name);
 557
 558         remove_file_hash(f);
 559
 560         if (close(f->fd) < 0)
 561                 ret = errno;
 562
 563         f->fd = -1;
 564
 565         if (f->shadow_fd != -1) {
 566                 close(f->shadow_fd);
 567                 f->shadow_fd = -1;
 568         }
 569
 570         f->engine_pos = 0;
 571         return ret;
 572 }
 573
 574 int file_lookup_open(struct fio_file *f, int flags)
 575 {
 576         struct fio_file *__f;
 577         int from_hash;
 578
 579         __f = lookup_file_hash(f->file_name);
 580         if (__f) {
 581                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 582                 f->lock = __f->lock;
 583                 from_hash = 1;
 584         } else {
 585                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 586                 from_hash = 0;
 587         }
 588
 589 #ifdef WIN32
 590         flags |= _O_BINARY;
 591 #endif
 592
 593         f->fd = open(f->file_name, flags, 0600);
 594         return from_hash;
 595 }
 596
 597 static int file_close_shadow_fds(struct thread_data *td)
 598 {
 599         struct fio_file *f;
 600         int num_closed = 0;
 601         unsigned int i;
 602
 603         for_each_file(td, f, i) {
 604                 if (f->shadow_fd == -1)
 605                         continue;
 606
 607                 close(f->shadow_fd);
 608                 f->shadow_fd = -1;
 609                 num_closed++;
 610         }
 611
 612         return num_closed;
 613 }
 614
 615 int generic_open_file(struct thread_data *td, struct fio_file *f)
 616 {
 617         int is_std = 0;
 618         int flags = 0;
 619         int from_hash = 0;
 620
 621         dprint(FD_FILE, "fd open %s\n", f->file_name);
 622
 623         if (!strcmp(f->file_name, "-")) {
 624                 if (td_rw(td)) {
 625                         log_err("fio: can't read/write to stdin/out\n");
 626                         return 1;
 627                 }
 628                 is_std = 1;
 629
 630                 /*
 631                  * move output logging to stderr, if we are writing to stdout
 632                  */
 633                 if (td_write(td))
 634                         f_out = stderr;
 635         }
 636
 637         if (td_trim(td))
 638                 goto skip_flags;
 639         if (td->o.odirect)
 640                 flags |= OS_O_DIRECT;
 641         if (td->o.oatomic) {
 642                 if (!FIO_O_ATOMIC) {
 643                         td_verror(td, EINVAL, "OS does not support atomic IO");
 644                         return 1;
 645                 }
 646                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 647         }
 648         if (td->o.sync_io)
 649                 flags |= O_SYNC;
 650         if (td->o.create_on_open && td->o.allow_create)
 651                 flags |= O_CREAT;
 652 skip_flags:
 653         if (f->filetype != FIO_TYPE_FILE)
 654                 flags |= FIO_O_NOATIME;
 655
 656 open_again:
 657         if (td_write(td)) {
 658                 if (!read_only)
 659                         flags |= O_RDWR;
 660
 661                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 662                         flags |= O_CREAT;
 663
 664                 if (is_std)
 665                         f->fd = dup(STDOUT_FILENO);
 666                 else
 667                         from_hash = file_lookup_open(f, flags);
 668         } else if (td_read(td)) {
 669                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 670                         flags |= O_RDWR;
 671                 else
 672                         flags |= O_RDONLY;
 673
 674                 if (is_std)
 675                         f->fd = dup(STDIN_FILENO);
 676                 else
 677                         from_hash = file_lookup_open(f, flags);
 678         } else if (td_trim(td)) {
 679                 assert(!td_rw(td)); /* should have matched above */
 680                 flags |= O_RDWR;
 681                 from_hash = file_lookup_open(f, flags);
 682         }
 683
 684         if (f->fd == -1) {
 685                 char buf[FIO_VERROR_SIZE];
 686                 int __e = errno;
 687
 688                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 689                         flags &= ~FIO_O_NOATIME;
 690                         goto open_again;
 691                 }
 692                 if (__e == EMFILE && file_close_shadow_fds(td))
 693                         goto open_again;
 694
 695                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 696
 697                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 698                         log_err("fio: looks like your file system does not " \
 699                                 "support direct=1/buffered=0\n");
 700                 }
 701
 702                 td_verror(td, __e, buf);
 703                 return 1;
 704         }
 705
 706         if (!from_hash && f->fd != -1) {
 707                 if (add_file_hash(f)) {
 708                         int fio_unused ret;
 709
 710                         /*
 711                          * Stash away descriptor for later close. This is to
 712                          * work-around a "feature" on Linux, where a close of
 713                          * an fd that has been opened for write will trigger
 714                          * udev to call blkid to check partitions, fs id, etc.
 715                          * That pollutes the device cache, which can slow down
 716                          * unbuffered accesses.
 717                          */
 718                         if (f->shadow_fd == -1)
 719                                 f->shadow_fd = f->fd;
 720                         else {
 721                                 /*
 722                                  * OK to ignore, we haven't done anything
 723                                  * with it
 724                                  */
 725                                 ret = generic_close_file(td, f);
 726                         }
 727                         goto open_again;
 728                 }
 729         }
 730
 731         return 0;
 732 }
 733
 734 /*
 735  * This function i.e. get_file_size() is the default .get_file_size
 736  * implementation of majority of I/O engines.
 737  */
 738 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 739 {
 740         return get_file_size(td, f);
 741 }
 742
 743 /*
 744  * open/close all files, so that ->real_file_size gets set
 745  */
 746 static int get_file_sizes(struct thread_data *td)
 747 {
 748         struct fio_file *f;
 749         unsigned int i;
 750         int err = 0;
 751
 752         for_each_file(td, f, i) {
 753                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 754                                                                 f->file_name);
 755
 756                 if (td_io_get_file_size(td, f)) {
 757                         if (td->error != ENOENT) {
 758                                 log_err("%s\n", td->verror);
 759                                 err = 1;
 760                                 break;
 761                         }
 762                         clear_error(td);
 763                 }
 764
 765                 /*
 766                  * There are corner cases where we end up with -1 for
 767                  * ->real_file_size due to unsupported file type, etc.
 768                  * We then just set to size option value divided by number
 769                  * of files, similar to the way file ->io_size is set.
 770                  * stat(2) failure doesn't set ->real_file_size to -1.
 771                  */
 772                 if (f->real_file_size == -1ULL && td->o.size)
 773                         f->real_file_size = td->o.size / td->o.nr_files;
 774         }
 775
 776         return err;
 777 }
 778
 779 struct fio_mount {
 780         struct flist_head list;
 781         const char *base;
 782         char __base[256];
 783         unsigned int key;
 784 };
 785
 786 /*
 787  * Get free number of bytes for each file on each unique mount.
 788  */
 789 static unsigned long long get_fs_free_counts(struct thread_data *td)
 790 {
 791         struct flist_head *n, *tmp;
 792         unsigned long long ret = 0;
 793         struct fio_mount *fm;
 794         FLIST_HEAD(list);
 795         struct fio_file *f;
 796         unsigned int i;
 797
 798         for_each_file(td, f, i) {
 799                 struct stat sb;
 800                 char buf[256];
 801
 802                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 803                         if (f->real_file_size != -1ULL)
 804                                 ret += f->real_file_size;
 805                         continue;
 806                 } else if (f->filetype != FIO_TYPE_FILE)
 807                         continue;
 808
 809                 buf[255] = '\0';
 810                 strncpy(buf, f->file_name, 255);
 811
 812                 if (stat(buf, &sb) < 0) {
 813                         if (errno != ENOENT)
 814                                 break;
 815                         strcpy(buf, ".");
 816                         if (stat(buf, &sb) < 0)
 817                                 break;
 818                 }
 819
 820                 fm = NULL;
 821                 flist_for_each(n, &list) {
 822                         fm = flist_entry(n, struct fio_mount, list);
 823                         if (fm->key == sb.st_dev)
 824                                 break;
 825
 826                         fm = NULL;
 827                 }
 828
 829                 if (fm)
 830                         continue;
 831
 832                 fm = calloc(1, sizeof(*fm));
 833                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 834                 fm->base = basename(fm->__base);
 835                 fm->key = sb.st_dev;
 836                 flist_add(&fm->list, &list);
 837         }
 838
 839         flist_for_each_safe(n, tmp, &list) {
 840                 unsigned long long sz;
 841
 842                 fm = flist_entry(n, struct fio_mount, list);
 843                 flist_del(&fm->list);
 844
 845                 sz = get_fs_free_size(fm->base);
 846                 if (sz && sz != -1ULL)
 847                         ret += sz;
 848
 849                 free(fm);
 850         }
 851
 852         return ret;
 853 }
 854
 855 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 856 {
 857         struct thread_options *o = &td->o;
 858         unsigned long long align_bs;
 859         unsigned long long offset;
 860
 861         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 862                 return f->real_file_size;
 863
 864         if (o->start_offset_percent > 0) {
 865                 /*
 866                  * if offset_align is provided, set initial offset
 867                  */
 868                 if (fio_option_is_set(o, start_offset_align)) {
 869                         align_bs = o->start_offset_align;
 870                 } else {
 871                         /* else take the minimum block size */
 872                         align_bs = td_min_bs(td);
 873                 }
 874
 875                 /* calculate the raw offset */
 876                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 877                         (td->subjob_number * o->offset_increment);
 878
 879                 /*
 880                  * block align the offset at the next available boundary at
 881                  * ceiling(offset / align_bs) * align_bs
 882                  */
 883                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 884
 885         } else {
 886                 /* start_offset_percent not set */
 887                 offset = o->start_offset +
 888                                 td->subjob_number * o->offset_increment;
 889         }
 890
 891         return offset;
 892 }
 893
 894 /*
 895  * Open the files and setup files sizes, creating files if necessary.
 896  */
 897 int setup_files(struct thread_data *td)
 898 {
 899         unsigned long long total_size, extend_size;
 900         struct thread_options *o = &td->o;
 901         struct fio_file *f;
 902         unsigned int i, nr_fs_extra = 0;
 903         int err = 0, need_extend;
 904         int old_state;
 905         const unsigned int bs = td_min_bs(td);
 906         uint64_t fs = 0;
 907
 908         dprint(FD_FILE, "setup files\n");
 909
 910         old_state = td_bump_runstate(td, TD_SETTING_UP);
 911
 912         if (o->read_iolog_file)
 913                 goto done;
 914
 915         /*
 916          * Find out physical size of files or devices for this thread,
 917          * before we determine I/O size and range of our targets.
 918          * If ioengine defines a setup() method, it's responsible for
 919          * opening the files and setting f->real_file_size to indicate
 920          * the valid range for that file.
 921          */
 922         if (td->io_ops->setup)
 923                 err = td->io_ops->setup(td);
 924         else
 925                 err = get_file_sizes(td);
 926
 927         if (err)
 928                 goto err_out;
 929
 930         /*
 931          * check sizes. if the files/devices do not exist and the size
 932          * isn't passed to fio, abort.
 933          */
 934         total_size = 0;
 935         for_each_file(td, f, i) {
 936                 f->fileno = i;
 937                 if (f->real_file_size == -1ULL)
 938                         total_size = -1ULL;
 939                 else
 940                         total_size += f->real_file_size;
 941         }
 942
 943         if (o->fill_device)
 944                 td->fill_device_size = get_fs_free_counts(td);
 945
 946         /*
 947          * device/file sizes are zero and no size given, punt
 948          */
 949         if ((!total_size || total_size == -1ULL) && !o->size &&
 950             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 951             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 952                 log_err("%s: you need to specify size=\n", o->name);
 953                 td_verror(td, EINVAL, "total_file_size");
 954                 goto err_out;
 955         }
 956
 957         /*
 958          * Calculate per-file size and potential extra size for the
 959          * first files, if needed (i.e. if we don't have a fixed size).
 960          */
 961         if (!o->file_size_low && o->nr_files) {
 962                 uint64_t all_fs;
 963
 964                 fs = o->size / o->nr_files;
 965                 all_fs = fs * o->nr_files;
 966
 967                 if (all_fs < o->size)
 968                         nr_fs_extra = (o->size - all_fs) / bs;
 969         }
 970
 971         /*
 972          * now file sizes are known, so we can set ->io_size. if size= is
 973          * not given, ->io_size is just equal to ->real_file_size. if size
 974          * is given, ->io_size is size / nr_files.
 975          */
 976         extend_size = total_size = 0;
 977         need_extend = 0;
 978         for_each_file(td, f, i) {
 979                 f->file_offset = get_start_offset(td, f);
 980
 981                 /*
 982                  * Update ->io_size depending on options specified.
 983                  * ->file_size_low being 0 means filesize option isn't set.
 984                  * Non zero ->file_size_low equals ->file_size_high means
 985                  * filesize option is set in a fixed size format.
 986                  * Non zero ->file_size_low not equals ->file_size_high means
 987                  * filesize option is set in a range format.
 988                  */
 989                 if (!o->file_size_low) {
 990                         /*
 991                          * no file size or range given, file size is equal to
 992                          * total size divided by number of files. If the size
 993                          * doesn't divide nicely with the min blocksize,
 994                          * make the first files bigger.
 995                          */
 996                         f->io_size = fs;
 997                         if (nr_fs_extra) {
 998                                 nr_fs_extra--;
 999                                 f->io_size += bs;
1000                         }
1001
1002                         /*
1003                          * We normally don't come here for regular files, but
1004                          * if the result is 0 for a regular file, set it to the
1005                          * real file size. This could be size of the existing
1006                          * one if it already exists, but otherwise will be set
1007                          * to 0. A new file won't be created because
1008                          * ->io_size + ->file_offset equals ->real_file_size.
1009                          */
1010                         if (!f->io_size) {
1011                                 if (f->file_offset > f->real_file_size)
1012                                         goto err_offset;
1013                                 f->io_size = f->real_file_size - f->file_offset;
1014                                 if (!f->io_size)
1015                                         log_info("fio: file %s may be ignored\n",
1016                                                 f->file_name);
1017                         }
1018                 } else if (f->real_file_size < o->file_size_low ||
1019                            f->real_file_size > o->file_size_high) {
1020                         if (f->file_offset > o->file_size_low)
1021                                 goto err_offset;
1022                         /*
1023                          * file size given. if it's fixed, use that. if it's a
1024                          * range, generate a random size in-between.
1025                          */
1026                         if (o->file_size_low == o->file_size_high)
1027                                 f->io_size = o->file_size_low - f->file_offset;
1028                         else {
1029                                 f->io_size = get_rand_file_size(td)
1030                                                 - f->file_offset;
1031                         }
1032                 } else
1033                         f->io_size = f->real_file_size - f->file_offset;
1034
1035                 if (f->io_size == -1ULL)
1036                         total_size = -1ULL;
1037                 else {
1038                         if (o->size_percent && o->size_percent != 100) {
1039                                 uint64_t file_size;
1040
1041                                 file_size = f->io_size + f->file_offset;
1042                                 f->io_size = (file_size *
1043                                               o->size_percent) / 100;
1044                                 if (f->io_size > (file_size - f->file_offset))
1045                                         f->io_size = file_size - f->file_offset;
1046
1047                                 f->io_size -= (f->io_size % td_min_bs(td));
1048                         }
1049                         total_size += f->io_size;
1050                 }
1051
1052                 if (f->filetype == FIO_TYPE_FILE &&
1053                     (f->io_size + f->file_offset) > f->real_file_size &&
1054                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1055                         if (!o->create_on_open) {
1056                                 need_extend++;
1057                                 extend_size += (f->io_size + f->file_offset);
1058                                 fio_file_set_extend(f);
1059                         } else
1060                                 f->real_file_size = f->io_size + f->file_offset;
1061                 }
1062         }
1063
1064         if (td->o.block_error_hist) {
1065                 int len;
1066
1067                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1068                 f = td->files[0];
1069                 len = f->io_size / td->o.bs[DDIR_TRIM];
1070                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1071                         log_err("fio: cannot calculate block histogram with "
1072                                 "%d trim blocks, maximum %d\n",
1073                                 len, MAX_NR_BLOCK_INFOS);
1074                         td_verror(td, EINVAL, "block_error_hist");
1075                         goto err_out;
1076                 }
1077
1078                 td->ts.nr_block_infos = len;
1079                 for (i = 0; i < len; i++)
1080                         td->ts.block_infos[i] =
1081                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1082         } else
1083                 td->ts.nr_block_infos = 0;
1084
1085         if (!o->size || (total_size && o->size > total_size))
1086                 o->size = total_size;
1087
1088         if (o->size < td_min_bs(td)) {
1089                 log_err("fio: blocksize too large for data set\n");
1090                 goto err_out;
1091         }
1092
1093         /*
1094          * See if we need to extend some files, typically needed when our
1095          * target regular files don't exist yet, but our jobs require them
1096          * initially due to read I/Os.
1097          */
1098         if (need_extend) {
1099                 temp_stall_ts = 1;
1100                 if (output_format & FIO_OUTPUT_NORMAL) {
1101                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1102                                  o->name,
1103                                  need_extend > 1 ? "s" : "",
1104                                  need_extend,
1105                                  need_extend > 1 ? "s" : "",
1106                                  need_extend > 1 ? "total " : "",
1107                                  extend_size >> 20);
1108                 }
1109
1110                 for_each_file(td, f, i) {
1111                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1112
1113                         if (!fio_file_extend(f))
1114                                 continue;
1115
1116                         assert(f->filetype == FIO_TYPE_FILE);
1117                         fio_file_clear_extend(f);
1118                         if (!o->fill_device) {
1119                                 old_len = f->real_file_size;
1120                                 extend_len = f->io_size + f->file_offset -
1121                                                 old_len;
1122                         }
1123                         f->real_file_size = (f->io_size + f->file_offset);
1124                         err = extend_file(td, f);
1125                         if (err)
1126                                 break;
1127
1128                         err = __file_invalidate_cache(td, f, old_len,
1129                                                                 extend_len);
1130
1131                         /*
1132                          * Shut up static checker
1133                          */
1134                         if (f->fd != -1)
1135                                 close(f->fd);
1136
1137                         f->fd = -1;
1138                         if (err)
1139                                 break;
1140                 }
1141                 temp_stall_ts = 0;
1142         }
1143
1144         if (err)
1145                 goto err_out;
1146
1147         if (!o->zone_size)
1148                 o->zone_size = o->size;
1149
1150         /*
1151          * iolog already set the total io size, if we read back
1152          * stored entries.
1153          */
1154         if (!o->read_iolog_file) {
1155                 if (o->io_size)
1156                         td->total_io_size = o->io_size * o->loops;
1157                 else
1158                         td->total_io_size = o->size * o->loops;
1159         }
1160
1161 done:
1162         if (o->create_only)
1163                 td->done = 1;
1164
1165         td_restore_runstate(td, old_state);
1166         return 0;
1167 err_offset:
1168         log_err("%s: you need to specify valid offset=\n", o->name);
1169 err_out:
1170         td_restore_runstate(td, old_state);
1171         return 1;
1172 }
1173
1174 bool pre_read_files(struct thread_data *td)
1175 {
1176         struct fio_file *f;
1177         unsigned int i;
1178
1179         dprint(FD_FILE, "pre_read files\n");
1180
1181         for_each_file(td, f, i) {
1182                 if (!pre_read_file(td, f))
1183                         return false;
1184         }
1185
1186         return true;
1187 }
1188
1189 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1190 {
1191         unsigned int range_size, seed;
1192         unsigned long nranges;
1193         uint64_t fsize;
1194
1195         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1196         fsize = min(f->real_file_size, f->io_size);
1197
1198         nranges = (fsize + range_size - 1) / range_size;
1199
1200         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1201         if (!td->o.rand_repeatable)
1202                 seed = td->rand_seeds[4];
1203
1204         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1205                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1206         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1207                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1208         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1209                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1210 }
1211
1212 static bool init_rand_distribution(struct thread_data *td)
1213 {
1214         struct fio_file *f;
1215         unsigned int i;
1216         int state;
1217
1218         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1219                 return false;
1220
1221         state = td_bump_runstate(td, TD_SETTING_UP);
1222
1223         for_each_file(td, f, i)
1224                 __init_rand_distribution(td, f);
1225
1226         td_restore_runstate(td, state);
1227         return true;
1228 }
1229
1230 /*
1231  * Check if the number of blocks exceeds the randomness capability of
1232  * the selected generator. Tausworthe is 32-bit, the others are fullly
1233  * 64-bit capable.
1234  */
1235 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1236                                  uint64_t blocks)
1237 {
1238         if (blocks <= FRAND32_MAX)
1239                 return 0;
1240         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1241                 return 0;
1242
1243         /*
1244          * If the user hasn't specified a random generator, switch
1245          * to tausworthe64 with informational warning. If the user did
1246          * specify one, just warn.
1247          */
1248         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1249                         f->file_name);
1250
1251         if (!fio_option_is_set(&td->o, random_generator)) {
1252                 log_info("fio: Switching to tausworthe64. Use the "
1253                          "random_generator= option to get rid of this "
1254                          "warning.\n");
1255                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1256                 return 0;
1257         }
1258
1259         /*
1260          * Just make this information to avoid breaking scripts.
1261          */
1262         log_info("fio: Use the random_generator= option to switch to lfsr or "
1263                          "tausworthe64.\n");
1264         return 0;
1265 }
1266
1267 bool init_random_map(struct thread_data *td)
1268 {
1269         unsigned long long blocks;
1270         struct fio_file *f;
1271         unsigned int i;
1272
1273         if (init_rand_distribution(td))
1274                 return true;
1275         if (!td_random(td))
1276                 return true;
1277
1278         for_each_file(td, f, i) {
1279                 uint64_t fsize = min(f->real_file_size, f->io_size);
1280
1281                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1282
1283                 if (check_rand_gen_limits(td, f, blocks))
1284                         return false;
1285
1286                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1287                         unsigned long seed;
1288
1289                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1290
1291                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1292                                 fio_file_set_lfsr(f);
1293                                 continue;
1294                         }
1295                 } else if (!td->o.norandommap) {
1296                         f->io_axmap = axmap_new(blocks);
1297                         if (f->io_axmap) {
1298                                 fio_file_set_axmap(f);
1299                                 continue;
1300                         }
1301                 } else if (td->o.norandommap)
1302                         continue;
1303
1304                 if (!td->o.softrandommap) {
1305                         log_err("fio: failed allocating random map. If running"
1306                                 " a large number of jobs, try the 'norandommap'"
1307                                 " option or set 'softrandommap'. Or give"
1308                                 " a larger --alloc-size to fio.\n");
1309                         return false;
1310                 }
1311
1312                 log_info("fio: file %s failed allocating random map. Running "
1313                          "job without.\n", f->file_name);
1314         }
1315
1316         return true;
1317 }
1318
1319 void close_files(struct thread_data *td)
1320 {
1321         struct fio_file *f;
1322         unsigned int i;
1323
1324         for_each_file(td, f, i) {
1325                 if (fio_file_open(f))
1326                         td_io_close_file(td, f);
1327         }
1328 }
1329
1330 void close_and_free_files(struct thread_data *td)
1331 {
1332         struct fio_file *f;
1333         unsigned int i;
1334         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1335
1336         dprint(FD_FILE, "close files\n");
1337
1338         for_each_file(td, f, i) {
1339                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1340                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1341                         td_io_unlink_file(td, f);
1342                 }
1343
1344                 if (fio_file_open(f))
1345                         td_io_close_file(td, f);
1346
1347                 remove_file_hash(f);
1348
1349                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1350                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1351                         td_io_unlink_file(td, f);
1352                 }
1353
1354                 if (use_free)
1355                         free(f->file_name);
1356                 else
1357                         sfree(f->file_name);
1358                 f->file_name = NULL;
1359                 if (fio_file_axmap(f)) {
1360                         axmap_free(f->io_axmap);
1361                         f->io_axmap = NULL;
1362                 }
1363                 if (use_free)
1364                         free(f);
1365                 else
1366                         sfree(f);
1367         }
1368
1369         td->o.filename = NULL;
1370         free(td->files);
1371         free(td->file_locks);
1372         td->files_index = 0;
1373         td->files = NULL;
1374         td->file_locks = NULL;
1375         td->o.file_lock_mode = FILE_LOCK_NONE;
1376         td->o.nr_files = 0;
1377 }
1378
1379 static void get_file_type(struct fio_file *f)
1380 {
1381         struct stat sb;
1382
1383         if (!strcmp(f->file_name, "-"))
1384                 f->filetype = FIO_TYPE_PIPE;
1385         else
1386                 f->filetype = FIO_TYPE_FILE;
1387
1388 #ifdef WIN32
1389         /* \\.\ is the device namespace in Windows, where every file is
1390          * a block device */
1391         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1392                 f->filetype = FIO_TYPE_BLOCK;
1393 #endif
1394
1395         if (!stat(f->file_name, &sb)) {
1396                 if (S_ISBLK(sb.st_mode))
1397                         f->filetype = FIO_TYPE_BLOCK;
1398                 else if (S_ISCHR(sb.st_mode))
1399                         f->filetype = FIO_TYPE_CHAR;
1400                 else if (S_ISFIFO(sb.st_mode))
1401                         f->filetype = FIO_TYPE_PIPE;
1402         }
1403 }
1404
1405 static bool __is_already_allocated(const char *fname, bool set)
1406 {
1407         struct flist_head *entry;
1408         bool ret;
1409
1410         ret = file_bloom_exists(fname, set);
1411         if (!ret)
1412                 return ret;
1413
1414         flist_for_each(entry, &filename_list) {
1415                 struct file_name *fn;
1416
1417                 fn = flist_entry(entry, struct file_name, list);
1418
1419                 if (!strcmp(fn->filename, fname))
1420                         return true;
1421         }
1422
1423         return false;
1424 }
1425
1426 static bool is_already_allocated(const char *fname)
1427 {
1428         bool ret;
1429
1430         fio_file_hash_lock();
1431         ret = __is_already_allocated(fname, false);
1432         fio_file_hash_unlock();
1433
1434         return ret;
1435 }
1436
1437 static void set_already_allocated(const char *fname)
1438 {
1439         struct file_name *fn;
1440
1441         fn = malloc(sizeof(struct file_name));
1442         fn->filename = strdup(fname);
1443
1444         fio_file_hash_lock();
1445         if (!__is_already_allocated(fname, true)) {
1446                 flist_add_tail(&fn->list, &filename_list);
1447                 fn = NULL;
1448         }
1449         fio_file_hash_unlock();
1450
1451         if (fn) {
1452                 free(fn->filename);
1453                 free(fn);
1454         }
1455 }
1456
1457 static void free_already_allocated(void)
1458 {
1459         struct flist_head *entry, *tmp;
1460         struct file_name *fn;
1461
1462         if (flist_empty(&filename_list))
1463                 return;
1464
1465         fio_file_hash_lock();
1466         flist_for_each_safe(entry, tmp, &filename_list) {
1467                 fn = flist_entry(entry, struct file_name, list);
1468                 free(fn->filename);
1469                 flist_del(&fn->list);
1470                 free(fn);
1471         }
1472
1473         fio_file_hash_unlock();
1474 }
1475
1476 static struct fio_file *alloc_new_file(struct thread_data *td)
1477 {
1478         struct fio_file *f;
1479
1480         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1481                 f = calloc(1, sizeof(*f));
1482         else
1483                 f = scalloc(1, sizeof(*f));
1484         if (!f) {
1485                 assert(0);
1486                 return NULL;
1487         }
1488
1489         f->fd = -1;
1490         f->shadow_fd = -1;
1491         fio_file_reset(td, f);
1492         return f;
1493 }
1494
1495 bool exists_and_not_regfile(const char *filename)
1496 {
1497         struct stat sb;
1498
1499         if (lstat(filename, &sb) == -1)
1500                 return false;
1501
1502 #ifndef WIN32 /* NOT Windows */
1503         if (S_ISREG(sb.st_mode))
1504                 return false;
1505 #else
1506         /* \\.\ is the device namespace in Windows, where every file
1507          * is a device node */
1508         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1509                 return false;
1510 #endif
1511
1512         return true;
1513 }
1514
1515 static bool create_work_dirs(struct thread_data *td, const char *fname)
1516 {
1517         char path[PATH_MAX];
1518         char *start, *end;
1519
1520         if (td->o.directory) {
1521                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1522                          FIO_OS_PATH_SEPARATOR, fname);
1523                 start = strstr(path, fname);
1524         } else {
1525                 snprintf(path, PATH_MAX, "%s", fname);
1526                 start = path;
1527         }
1528
1529         end = start;
1530         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1531                 if (end == start)
1532                         break;
1533                 *end = '\0';
1534                 errno = 0;
1535 #ifdef CONFIG_HAVE_MKDIR_TWO
1536                 if (mkdir(path, 0600) && errno != EEXIST) {
1537 #else
1538                 if (mkdir(path) && errno != EEXIST) {
1539 #endif
1540                         log_err("fio: failed to create dir (%s): %d\n",
1541                                 start, errno);
1542                         return false;
1543                 }
1544                 *end = FIO_OS_PATH_SEPARATOR;
1545                 end++;
1546         }
1547         td->flags |= TD_F_DIRS_CREATED;
1548         return true;
1549 }
1550
1551 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1552 {
1553         int cur_files = td->files_index;
1554         char file_name[PATH_MAX];
1555         struct fio_file *f;
1556         int len = 0;
1557
1558         dprint(FD_FILE, "add file %s\n", fname);
1559
1560         if (td->o.directory)
1561                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1562                                         td->o.unique_filename);
1563
1564         sprintf(file_name + len, "%s", fname);
1565
1566         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1567             !(td->flags & TD_F_DIRS_CREATED) &&
1568             !create_work_dirs(td, fname))
1569                 return 1;
1570
1571         /* clean cloned siblings using existing files */
1572         if (numjob && is_already_allocated(file_name) &&
1573             !exists_and_not_regfile(fname))
1574                 return 0;
1575
1576         f = alloc_new_file(td);
1577
1578         if (td->files_size <= td->files_index) {
1579                 unsigned int new_size = td->o.nr_files + 1;
1580
1581                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1582
1583                 td->files = realloc(td->files, new_size * sizeof(f));
1584                 if (td->files == NULL) {
1585                         log_err("fio: realloc OOM\n");
1586                         assert(0);
1587                 }
1588                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1589                         td->file_locks = realloc(td->file_locks, new_size);
1590                         if (!td->file_locks) {
1591                                 log_err("fio: realloc OOM\n");
1592                                 assert(0);
1593                         }
1594                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1595                 }
1596                 td->files_size = new_size;
1597         }
1598         td->files[cur_files] = f;
1599         f->fileno = cur_files;
1600
1601         /*
1602          * init function, io engine may not be loaded yet
1603          */
1604         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1605                 f->real_file_size = -1ULL;
1606
1607         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1608                 f->file_name = strdup(file_name);
1609         else
1610                 f->file_name = smalloc_strdup(file_name);
1611
1612         /* can't handle smalloc failure from here */
1613         assert(f->file_name);
1614
1615         get_file_type(f);
1616
1617         switch (td->o.file_lock_mode) {
1618         case FILE_LOCK_NONE:
1619                 break;
1620         case FILE_LOCK_READWRITE:
1621                 f->rwlock = fio_rwlock_init();
1622                 break;
1623         case FILE_LOCK_EXCLUSIVE:
1624                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1625                 break;
1626         default:
1627                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1628                 assert(0);
1629         }
1630
1631         td->files_index++;
1632
1633         if (td->o.numjobs > 1)
1634                 set_already_allocated(file_name);
1635
1636         if (inc)
1637                 td->o.nr_files++;
1638
1639         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1640                                                         cur_files);
1641
1642         return cur_files;
1643 }
1644
1645 int add_file_exclusive(struct thread_data *td, const char *fname)
1646 {
1647         struct fio_file *f;
1648         unsigned int i;
1649
1650         for_each_file(td, f, i) {
1651                 if (!strcmp(f->file_name, fname))
1652                         return i;
1653         }
1654
1655         return add_file(td, fname, 0, 1);
1656 }
1657
1658 void get_file(struct fio_file *f)
1659 {
1660         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1661         assert(fio_file_open(f));
1662         f->references++;
1663 }
1664
1665 int put_file(struct thread_data *td, struct fio_file *f)
1666 {
1667         int f_ret = 0, ret = 0;
1668
1669         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1670
1671         if (!fio_file_open(f)) {
1672                 assert(f->fd == -1);
1673                 return 0;
1674         }
1675
1676         assert(f->references);
1677         if (--f->references)
1678                 return 0;
1679
1680         if (should_fsync(td) && td->o.fsync_on_close) {
1681                 f_ret = fsync(f->fd);
1682                 if (f_ret < 0)
1683                         f_ret = errno;
1684         }
1685
1686         if (td->io_ops->close_file)
1687                 ret = td->io_ops->close_file(td, f);
1688
1689         if (!ret)
1690                 ret = f_ret;
1691
1692         td->nr_open_files--;
1693         fio_file_clear_open(f);
1694         assert(f->fd == -1);
1695         return ret;
1696 }
1697
1698 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1699 {
1700         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1701                 return;
1702
1703         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1704                 if (ddir == DDIR_READ)
1705                         fio_rwlock_read(f->rwlock);
1706                 else
1707                         fio_rwlock_write(f->rwlock);
1708         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1709                 fio_sem_down(f->lock);
1710
1711         td->file_locks[f->fileno] = td->o.file_lock_mode;
1712 }
1713
1714 void unlock_file(struct thread_data *td, struct fio_file *f)
1715 {
1716         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1717                 return;
1718
1719         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1720                 fio_rwlock_unlock(f->rwlock);
1721         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1722                 fio_sem_up(f->lock);
1723
1724         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1725 }
1726
1727 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1728 {
1729         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1730                 return;
1731         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1732                 unlock_file(td, f);
1733 }
1734
1735 static bool recurse_dir(struct thread_data *td, const char *dirname)
1736 {
1737         struct dirent *dir;
1738         bool ret = false;
1739         DIR *D;
1740
1741         D = opendir(dirname);
1742         if (!D) {
1743                 char buf[FIO_VERROR_SIZE];
1744
1745                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1746                 td_verror(td, errno, buf);
1747                 return true;
1748         }
1749
1750         while ((dir = readdir(D)) != NULL) {
1751                 char full_path[PATH_MAX];
1752                 struct stat sb;
1753
1754                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1755                         continue;
1756
1757                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1758
1759                 if (lstat(full_path, &sb) == -1) {
1760                         if (errno != ENOENT) {
1761                                 td_verror(td, errno, "stat");
1762                                 ret = true;
1763                                 break;
1764                         }
1765                 }
1766
1767                 if (S_ISREG(sb.st_mode)) {
1768                         add_file(td, full_path, 0, 1);
1769                         continue;
1770                 }
1771                 if (!S_ISDIR(sb.st_mode))
1772                         continue;
1773
1774                 ret = recurse_dir(td, full_path);
1775                 if (ret)
1776                         break;
1777         }
1778
1779         closedir(D);
1780         return ret;
1781 }
1782
1783 int add_dir_files(struct thread_data *td, const char *path)
1784 {
1785         int ret = recurse_dir(td, path);
1786
1787         if (!ret)
1788                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1789
1790         return ret;
1791 }
1792
1793 void dup_files(struct thread_data *td, struct thread_data *org)
1794 {
1795         struct fio_file *f;
1796         unsigned int i;
1797
1798         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1799
1800         if (!org->files)
1801                 return;
1802
1803         td->files = malloc(org->files_index * sizeof(f));
1804
1805         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1806                 td->file_locks = malloc(org->files_index);
1807
1808         for_each_file(org, f, i) {
1809                 struct fio_file *__f;
1810
1811                 __f = alloc_new_file(td);
1812
1813                 if (f->file_name) {
1814                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1815                                 __f->file_name = strdup(f->file_name);
1816                         else
1817                                 __f->file_name = smalloc_strdup(f->file_name);
1818
1819                         /* can't handle smalloc failure from here */
1820                         assert(__f->file_name);
1821                         __f->filetype = f->filetype;
1822                 }
1823
1824                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1825                         __f->lock = f->lock;
1826                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1827                         __f->rwlock = f->rwlock;
1828
1829                 td->files[i] = __f;
1830         }
1831 }
1832
1833 /*
1834  * Returns the index that matches the filename, or -1 if not there
1835  */
1836 int get_fileno(struct thread_data *td, const char *fname)
1837 {
1838         struct fio_file *f;
1839         unsigned int i;
1840
1841         for_each_file(td, f, i)
1842                 if (!strcmp(f->file_name, fname))
1843                         return i;
1844
1845         return -1;
1846 }
1847
1848 /*
1849  * For log usage, where we add/open/close files automatically
1850  */
1851 void free_release_files(struct thread_data *td)
1852 {
1853         close_files(td);
1854         td->o.nr_files = 0;
1855         td->o.open_files = 0;
1856         td->files_index = 0;
1857 }
1858
1859 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1860 {
1861         int i;
1862
1863         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1864                 f->last_pos[i] = f->file_offset;
1865                 f->last_start[i] = -1ULL;
1866         }
1867
1868         if (fio_file_axmap(f))
1869                 axmap_reset(f->io_axmap);
1870         else if (fio_file_lfsr(f))
1871                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1872 }
1873
1874 bool fio_files_done(struct thread_data *td)
1875 {
1876         struct fio_file *f;
1877         unsigned int i;
1878
1879         for_each_file(td, f, i)
1880                 if (!fio_file_done(f))
1881                         return false;
1882
1883         return true;
1884 }
1885
1886 /* free memory used in initialization phase only */
1887 void filesetup_mem_free(void)
1888 {
1889         free_already_allocated();
1890 }
1891
1892 /*
1893  * This function is for platforms which support direct I/O but not O_DIRECT.
1894  */
1895 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1896 {
1897 #ifdef FIO_OS_DIRECTIO
1898         int ret = fio_set_odirect(f);
1899
1900         if (ret) {
1901                 td_verror(td, ret, "fio_set_directio");
1902 #if defined(__sun__)
1903                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1904                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1905                 } else {
1906                         log_err("fio: the file system does not seem to support direct IO\n");
1907                 }
1908 #else
1909                 log_err("fio: the file system does not seem to support direct IO\n");
1910 #endif
1911                 return -1;
1912         }
1913
1914         return 0;
1915 #else
1916         log_err("fio: direct IO is not supported on this host operating system\n");
1917         return -1;
1918 #endif
1919 }