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