filesetup.c

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