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                 if (!read_only)
 678                         flags |= O_RDWR;
 679                 from_hash = file_lookup_open(f, flags);
 680         }
 681
 682         if (f->fd == -1) {
 683                 char buf[FIO_VERROR_SIZE];
 684                 int __e = errno;
 685
 686                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 687                         flags &= ~FIO_O_NOATIME;
 688                         goto open_again;
 689                 }
 690                 if (__e == EMFILE && file_close_shadow_fds(td))
 691                         goto open_again;
 692
 693                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 694
 695                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 696                         log_err("fio: looks like your file system does not " \
 697                                 "support direct=1/buffered=0\n");
 698                 }
 699
 700                 td_verror(td, __e, buf);
 701                 return 1;
 702         }
 703
 704         if (!from_hash && f->fd != -1) {
 705                 if (add_file_hash(f)) {
 706                         int fio_unused ret;
 707
 708                         /*
 709                          * Stash away descriptor for later close. This is to
 710                          * work-around a "feature" on Linux, where a close of
 711                          * an fd that has been opened for write will trigger
 712                          * udev to call blkid to check partitions, fs id, etc.
 713                          * That pollutes the device cache, which can slow down
 714                          * unbuffered accesses.
 715                          */
 716                         if (f->shadow_fd == -1)
 717                                 f->shadow_fd = f->fd;
 718                         else {
 719                                 /*
 720                                  * OK to ignore, we haven't done anything
 721                                  * with it
 722                                  */
 723                                 ret = generic_close_file(td, f);
 724                         }
 725                         goto open_again;
 726                 }
 727         }
 728
 729         return 0;
 730 }
 731
 732 /*
 733  * This function i.e. get_file_size() is the default .get_file_size
 734  * implementation of majority of I/O engines.
 735  */
 736 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 737 {
 738         return get_file_size(td, f);
 739 }
 740
 741 /*
 742  * open/close all files, so that ->real_file_size gets set
 743  */
 744 static int get_file_sizes(struct thread_data *td)
 745 {
 746         struct fio_file *f;
 747         unsigned int i;
 748         int err = 0;
 749
 750         for_each_file(td, f, i) {
 751                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 752                                                                 f->file_name);
 753
 754                 if (td_io_get_file_size(td, f)) {
 755                         if (td->error != ENOENT) {
 756                                 log_err("%s\n", td->verror);
 757                                 err = 1;
 758                                 break;
 759                         }
 760                         clear_error(td);
 761                 }
 762
 763                 /*
 764                  * There are corner cases where we end up with -1 for
 765                  * ->real_file_size due to unsupported file type, etc.
 766                  * We then just set to size option value divided by number
 767                  * of files, similar to the way file ->io_size is set.
 768                  * stat(2) failure doesn't set ->real_file_size to -1.
 769                  */
 770                 if (f->real_file_size == -1ULL && td->o.size)
 771                         f->real_file_size = td->o.size / td->o.nr_files;
 772         }
 773
 774         return err;
 775 }
 776
 777 struct fio_mount {
 778         struct flist_head list;
 779         const char *base;
 780         char __base[256];
 781         unsigned int key;
 782 };
 783
 784 /*
 785  * Get free number of bytes for each file on each unique mount.
 786  */
 787 static unsigned long long get_fs_free_counts(struct thread_data *td)
 788 {
 789         struct flist_head *n, *tmp;
 790         unsigned long long ret = 0;
 791         struct fio_mount *fm;
 792         FLIST_HEAD(list);
 793         struct fio_file *f;
 794         unsigned int i;
 795
 796         for_each_file(td, f, i) {
 797                 struct stat sb;
 798                 char buf[256];
 799
 800                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 801                         if (f->real_file_size != -1ULL)
 802                                 ret += f->real_file_size;
 803                         continue;
 804                 } else if (f->filetype != FIO_TYPE_FILE)
 805                         continue;
 806
 807                 buf[255] = '\0';
 808                 strncpy(buf, f->file_name, 255);
 809
 810                 if (stat(buf, &sb) < 0) {
 811                         if (errno != ENOENT)
 812                                 break;
 813                         strcpy(buf, ".");
 814                         if (stat(buf, &sb) < 0)
 815                                 break;
 816                 }
 817
 818                 fm = NULL;
 819                 flist_for_each(n, &list) {
 820                         fm = flist_entry(n, struct fio_mount, list);
 821                         if (fm->key == sb.st_dev)
 822                                 break;
 823
 824                         fm = NULL;
 825                 }
 826
 827                 if (fm)
 828                         continue;
 829
 830                 fm = calloc(1, sizeof(*fm));
 831                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 832                 fm->base = basename(fm->__base);
 833                 fm->key = sb.st_dev;
 834                 flist_add(&fm->list, &list);
 835         }
 836
 837         flist_for_each_safe(n, tmp, &list) {
 838                 unsigned long long sz;
 839
 840                 fm = flist_entry(n, struct fio_mount, list);
 841                 flist_del(&fm->list);
 842
 843                 sz = get_fs_free_size(fm->base);
 844                 if (sz && sz != -1ULL)
 845                         ret += sz;
 846
 847                 free(fm);
 848         }
 849
 850         return ret;
 851 }
 852
 853 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 854 {
 855         struct thread_options *o = &td->o;
 856         unsigned long long align_bs;
 857         unsigned long long offset;
 858
 859         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 860                 return f->real_file_size;
 861
 862         if (o->start_offset_percent > 0) {
 863                 /*
 864                  * if offset_align is provided, set initial offset
 865                  */
 866                 if (fio_option_is_set(o, start_offset_align)) {
 867                         align_bs = o->start_offset_align;
 868                 } else {
 869                         /* else take the minimum block size */
 870                         align_bs = td_min_bs(td);
 871                 }
 872
 873                 /* calculate the raw offset */
 874                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 875                         (td->subjob_number * o->offset_increment);
 876
 877                 /*
 878                  * block align the offset at the next available boundary at
 879                  * ceiling(offset / align_bs) * align_bs
 880                  */
 881                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 882
 883         } else {
 884                 /* start_offset_percent not set */
 885                 offset = o->start_offset +
 886                                 td->subjob_number * o->offset_increment;
 887         }
 888
 889         return offset;
 890 }
 891
 892 /*
 893  * Open the files and setup files sizes, creating files if necessary.
 894  */
 895 int setup_files(struct thread_data *td)
 896 {
 897         unsigned long long total_size, extend_size;
 898         struct thread_options *o = &td->o;
 899         struct fio_file *f;
 900         unsigned int i, nr_fs_extra = 0;
 901         int err = 0, need_extend;
 902         int old_state;
 903         const unsigned int bs = td_min_bs(td);
 904         uint64_t fs = 0;
 905
 906         dprint(FD_FILE, "setup files\n");
 907
 908         old_state = td_bump_runstate(td, TD_SETTING_UP);
 909
 910         if (o->read_iolog_file)
 911                 goto done;
 912
 913         /*
 914          * Find out physical size of files or devices for this thread,
 915          * before we determine I/O size and range of our targets.
 916          * If ioengine defines a setup() method, it's responsible for
 917          * opening the files and setting f->real_file_size to indicate
 918          * the valid range for that file.
 919          */
 920         if (td->io_ops->setup)
 921                 err = td->io_ops->setup(td);
 922         else
 923                 err = get_file_sizes(td);
 924
 925         if (err)
 926                 goto err_out;
 927
 928         /*
 929          * check sizes. if the files/devices do not exist and the size
 930          * isn't passed to fio, abort.
 931          */
 932         total_size = 0;
 933         for_each_file(td, f, i) {
 934                 f->fileno = i;
 935                 if (f->real_file_size == -1ULL)
 936                         total_size = -1ULL;
 937                 else
 938                         total_size += f->real_file_size;
 939         }
 940
 941         if (o->fill_device)
 942                 td->fill_device_size = get_fs_free_counts(td);
 943
 944         /*
 945          * device/file sizes are zero and no size given, punt
 946          */
 947         if ((!total_size || total_size == -1ULL) && !o->size &&
 948             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 949             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 950                 log_err("%s: you need to specify size=\n", o->name);
 951                 td_verror(td, EINVAL, "total_file_size");
 952                 goto err_out;
 953         }
 954
 955         /*
 956          * Calculate per-file size and potential extra size for the
 957          * first files, if needed (i.e. if we don't have a fixed size).
 958          */
 959         if (!o->file_size_low && o->nr_files) {
 960                 uint64_t all_fs;
 961
 962                 fs = o->size / o->nr_files;
 963                 all_fs = fs * o->nr_files;
 964
 965                 if (all_fs < o->size)
 966                         nr_fs_extra = (o->size - all_fs) / bs;
 967         }
 968
 969         /*
 970          * now file sizes are known, so we can set ->io_size. if size= is
 971          * not given, ->io_size is just equal to ->real_file_size. if size
 972          * is given, ->io_size is size / nr_files.
 973          */
 974         extend_size = total_size = 0;
 975         need_extend = 0;
 976         for_each_file(td, f, i) {
 977                 f->file_offset = get_start_offset(td, f);
 978
 979                 /*
 980                  * Update ->io_size depending on options specified.
 981                  * ->file_size_low being 0 means filesize option isn't set.
 982                  * Non zero ->file_size_low equals ->file_size_high means
 983                  * filesize option is set in a fixed size format.
 984                  * Non zero ->file_size_low not equals ->file_size_high means
 985                  * filesize option is set in a range format.
 986                  */
 987                 if (!o->file_size_low) {
 988                         /*
 989                          * no file size or range given, file size is equal to
 990                          * total size divided by number of files. If the size
 991                          * doesn't divide nicely with the min blocksize,
 992                          * make the first files bigger.
 993                          */
 994                         f->io_size = fs;
 995                         if (nr_fs_extra) {
 996                                 nr_fs_extra--;
 997                                 f->io_size += bs;
 998                         }
 999
1000                         /*
1001                          * We normally don't come here for regular files, but
1002                          * if the result is 0 for a regular file, set it to the
1003                          * real file size. This could be size of the existing
1004                          * one if it already exists, but otherwise will be set
1005                          * to 0. A new file won't be created because
1006                          * ->io_size + ->file_offset equals ->real_file_size.
1007                          */
1008                         if (!f->io_size) {
1009                                 if (f->file_offset > f->real_file_size)
1010                                         goto err_offset;
1011                                 f->io_size = f->real_file_size - f->file_offset;
1012                                 if (!f->io_size)
1013                                         log_info("fio: file %s may be ignored\n",
1014                                                 f->file_name);
1015                         }
1016                 } else if (f->real_file_size < o->file_size_low ||
1017                            f->real_file_size > o->file_size_high) {
1018                         if (f->file_offset > o->file_size_low)
1019                                 goto err_offset;
1020                         /*
1021                          * file size given. if it's fixed, use that. if it's a
1022                          * range, generate a random size in-between.
1023                          */
1024                         if (o->file_size_low == o->file_size_high)
1025                                 f->io_size = o->file_size_low - f->file_offset;
1026                         else {
1027                                 f->io_size = get_rand_file_size(td)
1028                                                 - f->file_offset;
1029                         }
1030                 } else
1031                         f->io_size = f->real_file_size - f->file_offset;
1032
1033                 if (f->io_size == -1ULL)
1034                         total_size = -1ULL;
1035                 else {
1036                         if (o->size_percent && o->size_percent != 100) {
1037                                 uint64_t file_size;
1038
1039                                 file_size = f->io_size + f->file_offset;
1040                                 f->io_size = (file_size *
1041                                               o->size_percent) / 100;
1042                                 if (f->io_size > (file_size - f->file_offset))
1043                                         f->io_size = file_size - f->file_offset;
1044
1045                                 f->io_size -= (f->io_size % td_min_bs(td));
1046                         }
1047                         total_size += f->io_size;
1048                 }
1049
1050                 if (f->filetype == FIO_TYPE_FILE &&
1051                     (f->io_size + f->file_offset) > f->real_file_size &&
1052                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1053                         if (!o->create_on_open) {
1054                                 need_extend++;
1055                                 extend_size += (f->io_size + f->file_offset);
1056                                 fio_file_set_extend(f);
1057                         } else
1058                                 f->real_file_size = f->io_size + f->file_offset;
1059                 }
1060         }
1061
1062         if (td->o.block_error_hist) {
1063                 int len;
1064
1065                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1066                 f = td->files[0];
1067                 len = f->io_size / td->o.bs[DDIR_TRIM];
1068                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1069                         log_err("fio: cannot calculate block histogram with "
1070                                 "%d trim blocks, maximum %d\n",
1071                                 len, MAX_NR_BLOCK_INFOS);
1072                         td_verror(td, EINVAL, "block_error_hist");
1073                         goto err_out;
1074                 }
1075
1076                 td->ts.nr_block_infos = len;
1077                 for (i = 0; i < len; i++)
1078                         td->ts.block_infos[i] =
1079                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1080         } else
1081                 td->ts.nr_block_infos = 0;
1082
1083         if (!o->size || (total_size && o->size > total_size))
1084                 o->size = total_size;
1085
1086         if (o->size < td_min_bs(td)) {
1087                 log_err("fio: blocksize too large for data set\n");
1088                 goto err_out;
1089         }
1090
1091         /*
1092          * See if we need to extend some files, typically needed when our
1093          * target regular files don't exist yet, but our jobs require them
1094          * initially due to read I/Os.
1095          */
1096         if (need_extend) {
1097                 temp_stall_ts = 1;
1098                 if (output_format & FIO_OUTPUT_NORMAL) {
1099                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1100                                  o->name,
1101                                  need_extend > 1 ? "s" : "",
1102                                  need_extend,
1103                                  need_extend > 1 ? "s" : "",
1104                                  need_extend > 1 ? "total " : "",
1105                                  extend_size >> 20);
1106                 }
1107
1108                 for_each_file(td, f, i) {
1109                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1110
1111                         if (!fio_file_extend(f))
1112                                 continue;
1113
1114                         assert(f->filetype == FIO_TYPE_FILE);
1115                         fio_file_clear_extend(f);
1116                         if (!o->fill_device) {
1117                                 old_len = f->real_file_size;
1118                                 extend_len = f->io_size + f->file_offset -
1119                                                 old_len;
1120                         }
1121                         f->real_file_size = (f->io_size + f->file_offset);
1122                         err = extend_file(td, f);
1123                         if (err)
1124                                 break;
1125
1126                         err = __file_invalidate_cache(td, f, old_len,
1127                                                                 extend_len);
1128
1129                         /*
1130                          * Shut up static checker
1131                          */
1132                         if (f->fd != -1)
1133                                 close(f->fd);
1134
1135                         f->fd = -1;
1136                         if (err)
1137                                 break;
1138                 }
1139                 temp_stall_ts = 0;
1140         }
1141
1142         if (err)
1143                 goto err_out;
1144
1145         if (!o->zone_size)
1146                 o->zone_size = o->size;
1147
1148         /*
1149          * iolog already set the total io size, if we read back
1150          * stored entries.
1151          */
1152         if (!o->read_iolog_file) {
1153                 if (o->io_size)
1154                         td->total_io_size = o->io_size * o->loops;
1155                 else
1156                         td->total_io_size = o->size * o->loops;
1157         }
1158
1159 done:
1160         if (o->create_only)
1161                 td->done = 1;
1162
1163         td_restore_runstate(td, old_state);
1164         return 0;
1165 err_offset:
1166         log_err("%s: you need to specify valid offset=\n", o->name);
1167 err_out:
1168         td_restore_runstate(td, old_state);
1169         return 1;
1170 }
1171
1172 bool pre_read_files(struct thread_data *td)
1173 {
1174         struct fio_file *f;
1175         unsigned int i;
1176
1177         dprint(FD_FILE, "pre_read files\n");
1178
1179         for_each_file(td, f, i) {
1180                 if (!pre_read_file(td, f))
1181                         return false;
1182         }
1183
1184         return true;
1185 }
1186
1187 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1188 {
1189         unsigned int range_size, seed;
1190         unsigned long nranges;
1191         uint64_t fsize;
1192
1193         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1194         fsize = min(f->real_file_size, f->io_size);
1195
1196         nranges = (fsize + range_size - 1) / range_size;
1197
1198         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1199         if (!td->o.rand_repeatable)
1200                 seed = td->rand_seeds[4];
1201
1202         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1203                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1204         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1205                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1206         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1207                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1208 }
1209
1210 static bool init_rand_distribution(struct thread_data *td)
1211 {
1212         struct fio_file *f;
1213         unsigned int i;
1214         int state;
1215
1216         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1217                 return false;
1218
1219         state = td_bump_runstate(td, TD_SETTING_UP);
1220
1221         for_each_file(td, f, i)
1222                 __init_rand_distribution(td, f);
1223
1224         td_restore_runstate(td, state);
1225         return true;
1226 }
1227
1228 /*
1229  * Check if the number of blocks exceeds the randomness capability of
1230  * the selected generator. Tausworthe is 32-bit, the others are fullly
1231  * 64-bit capable.
1232  */
1233 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1234                                  uint64_t blocks)
1235 {
1236         if (blocks <= FRAND32_MAX)
1237                 return 0;
1238         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1239                 return 0;
1240
1241         /*
1242          * If the user hasn't specified a random generator, switch
1243          * to tausworthe64 with informational warning. If the user did
1244          * specify one, just warn.
1245          */
1246         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1247                         f->file_name);
1248
1249         if (!fio_option_is_set(&td->o, random_generator)) {
1250                 log_info("fio: Switching to tausworthe64. Use the "
1251                          "random_generator= option to get rid of this "
1252                          "warning.\n");
1253                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1254                 return 0;
1255         }
1256
1257         /*
1258          * Just make this information to avoid breaking scripts.
1259          */
1260         log_info("fio: Use the random_generator= option to switch to lfsr or "
1261                          "tausworthe64.\n");
1262         return 0;
1263 }
1264
1265 bool init_random_map(struct thread_data *td)
1266 {
1267         unsigned long long blocks;
1268         struct fio_file *f;
1269         unsigned int i;
1270
1271         if (init_rand_distribution(td))
1272                 return true;
1273         if (!td_random(td))
1274                 return true;
1275
1276         for_each_file(td, f, i) {
1277                 uint64_t fsize = min(f->real_file_size, f->io_size);
1278
1279                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1280
1281                 if (check_rand_gen_limits(td, f, blocks))
1282                         return false;
1283
1284                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1285                         unsigned long seed;
1286
1287                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1288
1289                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1290                                 fio_file_set_lfsr(f);
1291                                 continue;
1292                         }
1293                 } else if (!td->o.norandommap) {
1294                         f->io_axmap = axmap_new(blocks);
1295                         if (f->io_axmap) {
1296                                 fio_file_set_axmap(f);
1297                                 continue;
1298                         }
1299                 } else if (td->o.norandommap)
1300                         continue;
1301
1302                 if (!td->o.softrandommap) {
1303                         log_err("fio: failed allocating random map. If running"
1304                                 " a large number of jobs, try the 'norandommap'"
1305                                 " option or set 'softrandommap'. Or give"
1306                                 " a larger --alloc-size to fio.\n");
1307                         return false;
1308                 }
1309
1310                 log_info("fio: file %s failed allocating random map. Running "
1311                          "job without.\n", f->file_name);
1312         }
1313
1314         return true;
1315 }
1316
1317 void close_files(struct thread_data *td)
1318 {
1319         struct fio_file *f;
1320         unsigned int i;
1321
1322         for_each_file(td, f, i) {
1323                 if (fio_file_open(f))
1324                         td_io_close_file(td, f);
1325         }
1326 }
1327
1328 void close_and_free_files(struct thread_data *td)
1329 {
1330         struct fio_file *f;
1331         unsigned int i;
1332         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1333
1334         dprint(FD_FILE, "close files\n");
1335
1336         for_each_file(td, f, i) {
1337                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1338                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1339                         td_io_unlink_file(td, f);
1340                 }
1341
1342                 if (fio_file_open(f))
1343                         td_io_close_file(td, f);
1344
1345                 remove_file_hash(f);
1346
1347                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1348                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1349                         td_io_unlink_file(td, f);
1350                 }
1351
1352                 if (use_free)
1353                         free(f->file_name);
1354                 else
1355                         sfree(f->file_name);
1356                 f->file_name = NULL;
1357                 if (fio_file_axmap(f)) {
1358                         axmap_free(f->io_axmap);
1359                         f->io_axmap = NULL;
1360                 }
1361                 if (use_free)
1362                         free(f);
1363                 else
1364                         sfree(f);
1365         }
1366
1367         td->o.filename = NULL;
1368         free(td->files);
1369         free(td->file_locks);
1370         td->files_index = 0;
1371         td->files = NULL;
1372         td->file_locks = NULL;
1373         td->o.file_lock_mode = FILE_LOCK_NONE;
1374         td->o.nr_files = 0;
1375 }
1376
1377 static void get_file_type(struct fio_file *f)
1378 {
1379         struct stat sb;
1380
1381         if (!strcmp(f->file_name, "-"))
1382                 f->filetype = FIO_TYPE_PIPE;
1383         else
1384                 f->filetype = FIO_TYPE_FILE;
1385
1386 #ifdef WIN32
1387         /* \\.\ is the device namespace in Windows, where every file is
1388          * a block device */
1389         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1390                 f->filetype = FIO_TYPE_BLOCK;
1391 #endif
1392
1393         if (!stat(f->file_name, &sb)) {
1394                 if (S_ISBLK(sb.st_mode))
1395                         f->filetype = FIO_TYPE_BLOCK;
1396                 else if (S_ISCHR(sb.st_mode))
1397                         f->filetype = FIO_TYPE_CHAR;
1398                 else if (S_ISFIFO(sb.st_mode))
1399                         f->filetype = FIO_TYPE_PIPE;
1400         }
1401 }
1402
1403 static bool __is_already_allocated(const char *fname, bool set)
1404 {
1405         struct flist_head *entry;
1406         bool ret;
1407
1408         ret = file_bloom_exists(fname, set);
1409         if (!ret)
1410                 return ret;
1411
1412         flist_for_each(entry, &filename_list) {
1413                 struct file_name *fn;
1414
1415                 fn = flist_entry(entry, struct file_name, list);
1416
1417                 if (!strcmp(fn->filename, fname))
1418                         return true;
1419         }
1420
1421         return false;
1422 }
1423
1424 static bool is_already_allocated(const char *fname)
1425 {
1426         bool ret;
1427
1428         fio_file_hash_lock();
1429         ret = __is_already_allocated(fname, false);
1430         fio_file_hash_unlock();
1431
1432         return ret;
1433 }
1434
1435 static void set_already_allocated(const char *fname)
1436 {
1437         struct file_name *fn;
1438
1439         fn = malloc(sizeof(struct file_name));
1440         fn->filename = strdup(fname);
1441
1442         fio_file_hash_lock();
1443         if (!__is_already_allocated(fname, true)) {
1444                 flist_add_tail(&fn->list, &filename_list);
1445                 fn = NULL;
1446         }
1447         fio_file_hash_unlock();
1448
1449         if (fn) {
1450                 free(fn->filename);
1451                 free(fn);
1452         }
1453 }
1454
1455 static void free_already_allocated(void)
1456 {
1457         struct flist_head *entry, *tmp;
1458         struct file_name *fn;
1459
1460         if (flist_empty(&filename_list))
1461                 return;
1462
1463         fio_file_hash_lock();
1464         flist_for_each_safe(entry, tmp, &filename_list) {
1465                 fn = flist_entry(entry, struct file_name, list);
1466                 free(fn->filename);
1467                 flist_del(&fn->list);
1468                 free(fn);
1469         }
1470
1471         fio_file_hash_unlock();
1472 }
1473
1474 static struct fio_file *alloc_new_file(struct thread_data *td)
1475 {
1476         struct fio_file *f;
1477
1478         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1479                 f = calloc(1, sizeof(*f));
1480         else
1481                 f = scalloc(1, sizeof(*f));
1482         if (!f) {
1483                 assert(0);
1484                 return NULL;
1485         }
1486
1487         f->fd = -1;
1488         f->shadow_fd = -1;
1489         fio_file_reset(td, f);
1490         return f;
1491 }
1492
1493 bool exists_and_not_regfile(const char *filename)
1494 {
1495         struct stat sb;
1496
1497         if (lstat(filename, &sb) == -1)
1498                 return false;
1499
1500 #ifndef WIN32 /* NOT Windows */
1501         if (S_ISREG(sb.st_mode))
1502                 return false;
1503 #else
1504         /* \\.\ is the device namespace in Windows, where every file
1505          * is a device node */
1506         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1507                 return false;
1508 #endif
1509
1510         return true;
1511 }
1512
1513 static bool create_work_dirs(struct thread_data *td, const char *fname)
1514 {
1515         char path[PATH_MAX];
1516         char *start, *end;
1517
1518         if (td->o.directory) {
1519                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1520                          FIO_OS_PATH_SEPARATOR, fname);
1521                 start = strstr(path, fname);
1522         } else {
1523                 snprintf(path, PATH_MAX, "%s", fname);
1524                 start = path;
1525         }
1526
1527         end = start;
1528         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1529                 if (end == start)
1530                         break;
1531                 *end = '\0';
1532                 errno = 0;
1533 #ifdef CONFIG_HAVE_MKDIR_TWO
1534                 if (mkdir(path, 0600) && errno != EEXIST) {
1535 #else
1536                 if (mkdir(path) && errno != EEXIST) {
1537 #endif
1538                         log_err("fio: failed to create dir (%s): %d\n",
1539                                 start, errno);
1540                         return false;
1541                 }
1542                 *end = FIO_OS_PATH_SEPARATOR;
1543                 end++;
1544         }
1545         td->flags |= TD_F_DIRS_CREATED;
1546         return true;
1547 }
1548
1549 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1550 {
1551         int cur_files = td->files_index;
1552         char file_name[PATH_MAX];
1553         struct fio_file *f;
1554         int len = 0;
1555
1556         dprint(FD_FILE, "add file %s\n", fname);
1557
1558         if (td->o.directory)
1559                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1560                                         td->o.unique_filename);
1561
1562         sprintf(file_name + len, "%s", fname);
1563
1564         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1565             !(td->flags & TD_F_DIRS_CREATED) &&
1566             !create_work_dirs(td, fname))
1567                 return 1;
1568
1569         /* clean cloned siblings using existing files */
1570         if (numjob && is_already_allocated(file_name) &&
1571             !exists_and_not_regfile(fname))
1572                 return 0;
1573
1574         f = alloc_new_file(td);
1575
1576         if (td->files_size <= td->files_index) {
1577                 unsigned int new_size = td->o.nr_files + 1;
1578
1579                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1580
1581                 td->files = realloc(td->files, new_size * sizeof(f));
1582                 if (td->files == NULL) {
1583                         log_err("fio: realloc OOM\n");
1584                         assert(0);
1585                 }
1586                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1587                         td->file_locks = realloc(td->file_locks, new_size);
1588                         if (!td->file_locks) {
1589                                 log_err("fio: realloc OOM\n");
1590                                 assert(0);
1591                         }
1592                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1593                 }
1594                 td->files_size = new_size;
1595         }
1596         td->files[cur_files] = f;
1597         f->fileno = cur_files;
1598
1599         /*
1600          * init function, io engine may not be loaded yet
1601          */
1602         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1603                 f->real_file_size = -1ULL;
1604
1605         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1606                 f->file_name = strdup(file_name);
1607         else
1608                 f->file_name = smalloc_strdup(file_name);
1609
1610         /* can't handle smalloc failure from here */
1611         assert(f->file_name);
1612
1613         get_file_type(f);
1614
1615         switch (td->o.file_lock_mode) {
1616         case FILE_LOCK_NONE:
1617                 break;
1618         case FILE_LOCK_READWRITE:
1619                 f->rwlock = fio_rwlock_init();
1620                 break;
1621         case FILE_LOCK_EXCLUSIVE:
1622                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1623                 break;
1624         default:
1625                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1626                 assert(0);
1627         }
1628
1629         td->files_index++;
1630
1631         if (td->o.numjobs > 1)
1632                 set_already_allocated(file_name);
1633
1634         if (inc)
1635                 td->o.nr_files++;
1636
1637         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1638                                                         cur_files);
1639
1640         return cur_files;
1641 }
1642
1643 int add_file_exclusive(struct thread_data *td, const char *fname)
1644 {
1645         struct fio_file *f;
1646         unsigned int i;
1647
1648         for_each_file(td, f, i) {
1649                 if (!strcmp(f->file_name, fname))
1650                         return i;
1651         }
1652
1653         return add_file(td, fname, 0, 1);
1654 }
1655
1656 void get_file(struct fio_file *f)
1657 {
1658         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1659         assert(fio_file_open(f));
1660         f->references++;
1661 }
1662
1663 int put_file(struct thread_data *td, struct fio_file *f)
1664 {
1665         int f_ret = 0, ret = 0;
1666
1667         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1668
1669         if (!fio_file_open(f)) {
1670                 assert(f->fd == -1);
1671                 return 0;
1672         }
1673
1674         assert(f->references);
1675         if (--f->references)
1676                 return 0;
1677
1678         if (should_fsync(td) && td->o.fsync_on_close) {
1679                 f_ret = fsync(f->fd);
1680                 if (f_ret < 0)
1681                         f_ret = errno;
1682         }
1683
1684         if (td->io_ops->close_file)
1685                 ret = td->io_ops->close_file(td, f);
1686
1687         if (!ret)
1688                 ret = f_ret;
1689
1690         td->nr_open_files--;
1691         fio_file_clear_open(f);
1692         assert(f->fd == -1);
1693         return ret;
1694 }
1695
1696 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1697 {
1698         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1699                 return;
1700
1701         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1702                 if (ddir == DDIR_READ)
1703                         fio_rwlock_read(f->rwlock);
1704                 else
1705                         fio_rwlock_write(f->rwlock);
1706         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1707                 fio_sem_down(f->lock);
1708
1709         td->file_locks[f->fileno] = td->o.file_lock_mode;
1710 }
1711
1712 void unlock_file(struct thread_data *td, struct fio_file *f)
1713 {
1714         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1715                 return;
1716
1717         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1718                 fio_rwlock_unlock(f->rwlock);
1719         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1720                 fio_sem_up(f->lock);
1721
1722         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1723 }
1724
1725 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1726 {
1727         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1728                 return;
1729         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1730                 unlock_file(td, f);
1731 }
1732
1733 static bool recurse_dir(struct thread_data *td, const char *dirname)
1734 {
1735         struct dirent *dir;
1736         bool ret = false;
1737         DIR *D;
1738
1739         D = opendir(dirname);
1740         if (!D) {
1741                 char buf[FIO_VERROR_SIZE];
1742
1743                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1744                 td_verror(td, errno, buf);
1745                 return true;
1746         }
1747
1748         while ((dir = readdir(D)) != NULL) {
1749                 char full_path[PATH_MAX];
1750                 struct stat sb;
1751
1752                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1753                         continue;
1754
1755                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1756
1757                 if (lstat(full_path, &sb) == -1) {
1758                         if (errno != ENOENT) {
1759                                 td_verror(td, errno, "stat");
1760                                 ret = true;
1761                                 break;
1762                         }
1763                 }
1764
1765                 if (S_ISREG(sb.st_mode)) {
1766                         add_file(td, full_path, 0, 1);
1767                         continue;
1768                 }
1769                 if (!S_ISDIR(sb.st_mode))
1770                         continue;
1771
1772                 ret = recurse_dir(td, full_path);
1773                 if (ret)
1774                         break;
1775         }
1776
1777         closedir(D);
1778         return ret;
1779 }
1780
1781 int add_dir_files(struct thread_data *td, const char *path)
1782 {
1783         int ret = recurse_dir(td, path);
1784
1785         if (!ret)
1786                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1787
1788         return ret;
1789 }
1790
1791 void dup_files(struct thread_data *td, struct thread_data *org)
1792 {
1793         struct fio_file *f;
1794         unsigned int i;
1795
1796         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1797
1798         if (!org->files)
1799                 return;
1800
1801         td->files = malloc(org->files_index * sizeof(f));
1802
1803         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1804                 td->file_locks = malloc(org->files_index);
1805
1806         for_each_file(org, f, i) {
1807                 struct fio_file *__f;
1808
1809                 __f = alloc_new_file(td);
1810
1811                 if (f->file_name) {
1812                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1813                                 __f->file_name = strdup(f->file_name);
1814                         else
1815                                 __f->file_name = smalloc_strdup(f->file_name);
1816
1817                         /* can't handle smalloc failure from here */
1818                         assert(__f->file_name);
1819                         __f->filetype = f->filetype;
1820                 }
1821
1822                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1823                         __f->lock = f->lock;
1824                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1825                         __f->rwlock = f->rwlock;
1826
1827                 td->files[i] = __f;
1828         }
1829 }
1830
1831 /*
1832  * Returns the index that matches the filename, or -1 if not there
1833  */
1834 int get_fileno(struct thread_data *td, const char *fname)
1835 {
1836         struct fio_file *f;
1837         unsigned int i;
1838
1839         for_each_file(td, f, i)
1840                 if (!strcmp(f->file_name, fname))
1841                         return i;
1842
1843         return -1;
1844 }
1845
1846 /*
1847  * For log usage, where we add/open/close files automatically
1848  */
1849 void free_release_files(struct thread_data *td)
1850 {
1851         close_files(td);
1852         td->o.nr_files = 0;
1853         td->o.open_files = 0;
1854         td->files_index = 0;
1855 }
1856
1857 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1858 {
1859         int i;
1860
1861         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1862                 f->last_pos[i] = f->file_offset;
1863                 f->last_start[i] = -1ULL;
1864         }
1865
1866         if (fio_file_axmap(f))
1867                 axmap_reset(f->io_axmap);
1868         else if (fio_file_lfsr(f))
1869                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1870 }
1871
1872 bool fio_files_done(struct thread_data *td)
1873 {
1874         struct fio_file *f;
1875         unsigned int i;
1876
1877         for_each_file(td, f, i)
1878                 if (!fio_file_done(f))
1879                         return false;
1880
1881         return true;
1882 }
1883
1884 /* free memory used in initialization phase only */
1885 void filesetup_mem_free(void)
1886 {
1887         free_already_allocated();
1888 }
1889
1890 /*
1891  * This function is for platforms which support direct I/O but not O_DIRECT.
1892  */
1893 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1894 {
1895 #ifdef FIO_OS_DIRECTIO
1896         int ret = fio_set_odirect(f);
1897
1898         if (ret) {
1899                 td_verror(td, ret, "fio_set_directio");
1900 #if defined(__sun__)
1901                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1902                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1903                 } else {
1904                         log_err("fio: the file system does not seem to support direct IO\n");
1905                 }
1906 #else
1907                 log_err("fio: the file system does not seem to support direct IO\n");
1908 #endif
1909                 return -1;
1910         }
1911
1912         return 0;
1913 #else
1914         log_err("fio: direct IO is not supported on this host operating system\n");
1915         return -1;
1916 #endif
1917 }