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 #include "zbd.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 long long 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 long long 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         } else if (f->filetype == FIO_TYPE_FILE) {
 495                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 496                         f->file_name, off, len);
 497                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 498                 if (ret)
 499                         errval = ret;
 500         } else if (f->filetype == FIO_TYPE_BLOCK) {
 501                 int retry_count = 0;
 502
 503                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 504                 ret = blockdev_invalidate_cache(f);
 505                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 506                         /*
 507                          * Linux multipath devices reject ioctl while
 508                          * the maps are being updated. That window can
 509                          * last tens of milliseconds; we'll try up to
 510                          * a quarter of a second.
 511                          */
 512                         usleep(10000);
 513                         ret = blockdev_invalidate_cache(f);
 514                 }
 515                 if (ret < 0 && errno == EACCES && geteuid()) {
 516                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 517                                 log_err("fio: only root may flush block "
 518                                         "devices. Cache flush bypassed!\n");
 519                         }
 520                 }
 521                 if (ret < 0)
 522                         errval = errno;
 523         } else if (f->filetype == FIO_TYPE_CHAR ||
 524                    f->filetype == FIO_TYPE_PIPE) {
 525                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 526         }
 527
 528         /*
 529          * Cache flushing isn't a fatal condition, and we know it will
 530          * happen on some platforms where we don't have the proper
 531          * function to flush eg block device caches. So just warn and
 532          * continue on our way.
 533          */
 534         if (errval)
 535                 log_info("fio: cache invalidation of %s failed: %s\n",
 536                          f->file_name, strerror(errval));
 537
 538         return 0;
 539
 540 }
 541
 542 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 543 {
 544         if (!fio_file_open(f))
 545                 return 0;
 546
 547         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 548 }
 549
 550 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 551 {
 552         int ret = 0;
 553
 554         dprint(FD_FILE, "fd close %s\n", f->file_name);
 555
 556         remove_file_hash(f);
 557
 558         if (close(f->fd) < 0)
 559                 ret = errno;
 560
 561         f->fd = -1;
 562
 563         if (f->shadow_fd != -1) {
 564                 close(f->shadow_fd);
 565                 f->shadow_fd = -1;
 566         }
 567
 568         f->engine_pos = 0;
 569         return ret;
 570 }
 571
 572 int file_lookup_open(struct fio_file *f, int flags)
 573 {
 574         struct fio_file *__f;
 575         int from_hash;
 576
 577         __f = lookup_file_hash(f->file_name);
 578         if (__f) {
 579                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 580                 f->lock = __f->lock;
 581                 from_hash = 1;
 582         } else {
 583                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 584                 from_hash = 0;
 585         }
 586
 587 #ifdef WIN32
 588         flags |= _O_BINARY;
 589 #endif
 590
 591         f->fd = open(f->file_name, flags, 0600);
 592         return from_hash;
 593 }
 594
 595 static int file_close_shadow_fds(struct thread_data *td)
 596 {
 597         struct fio_file *f;
 598         int num_closed = 0;
 599         unsigned int i;
 600
 601         for_each_file(td, f, i) {
 602                 if (f->shadow_fd == -1)
 603                         continue;
 604
 605                 close(f->shadow_fd);
 606                 f->shadow_fd = -1;
 607                 num_closed++;
 608         }
 609
 610         return num_closed;
 611 }
 612
 613 int generic_open_file(struct thread_data *td, struct fio_file *f)
 614 {
 615         int is_std = 0;
 616         int flags = 0;
 617         int from_hash = 0;
 618
 619         dprint(FD_FILE, "fd open %s\n", f->file_name);
 620
 621         if (!strcmp(f->file_name, "-")) {
 622                 if (td_rw(td)) {
 623                         log_err("fio: can't read/write to stdin/out\n");
 624                         return 1;
 625                 }
 626                 is_std = 1;
 627
 628                 /*
 629                  * move output logging to stderr, if we are writing to stdout
 630                  */
 631                 if (td_write(td))
 632                         f_out = stderr;
 633         }
 634
 635         if (td_trim(td))
 636                 goto skip_flags;
 637         if (td->o.odirect)
 638                 flags |= OS_O_DIRECT;
 639         if (td->o.oatomic) {
 640                 if (!FIO_O_ATOMIC) {
 641                         td_verror(td, EINVAL, "OS does not support atomic IO");
 642                         return 1;
 643                 }
 644                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 645         }
 646         if (td->o.sync_io)
 647                 flags |= O_SYNC;
 648         if (td->o.create_on_open && td->o.allow_create)
 649                 flags |= O_CREAT;
 650 skip_flags:
 651         if (f->filetype != FIO_TYPE_FILE)
 652                 flags |= FIO_O_NOATIME;
 653
 654 open_again:
 655         if (td_write(td)) {
 656                 if (!read_only)
 657                         flags |= O_RDWR;
 658
 659                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 660                         flags |= O_CREAT;
 661
 662                 if (is_std)
 663                         f->fd = dup(STDOUT_FILENO);
 664                 else
 665                         from_hash = file_lookup_open(f, flags);
 666         } else if (td_read(td)) {
 667                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 668                         flags |= O_RDWR;
 669                 else
 670                         flags |= O_RDONLY;
 671
 672                 if (is_std)
 673                         f->fd = dup(STDIN_FILENO);
 674                 else
 675                         from_hash = file_lookup_open(f, flags);
 676         } else if (td_trim(td)) {
 677                 assert(!td_rw(td)); /* should have matched above */
 678                 if (!read_only)
 679                         flags |= O_RDWR;
 680                 from_hash = file_lookup_open(f, flags);
 681         }
 682
 683         if (f->fd == -1) {
 684                 char buf[FIO_VERROR_SIZE];
 685                 int __e = errno;
 686
 687                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 688                         flags &= ~FIO_O_NOATIME;
 689                         goto open_again;
 690                 }
 691                 if (__e == EMFILE && file_close_shadow_fds(td))
 692                         goto open_again;
 693
 694                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 695
 696                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 697                         log_err("fio: looks like your file system does not " \
 698                                 "support direct=1/buffered=0\n");
 699                 }
 700
 701                 td_verror(td, __e, buf);
 702                 return 1;
 703         }
 704
 705         if (!from_hash && f->fd != -1) {
 706                 if (add_file_hash(f)) {
 707                         int fio_unused ret;
 708
 709                         /*
 710                          * Stash away descriptor for later close. This is to
 711                          * work-around a "feature" on Linux, where a close of
 712                          * an fd that has been opened for write will trigger
 713                          * udev to call blkid to check partitions, fs id, etc.
 714                          * That pollutes the device cache, which can slow down
 715                          * unbuffered accesses.
 716                          */
 717                         if (f->shadow_fd == -1)
 718                                 f->shadow_fd = f->fd;
 719                         else {
 720                                 /*
 721                                  * OK to ignore, we haven't done anything
 722                                  * with it
 723                                  */
 724                                 ret = generic_close_file(td, f);
 725                         }
 726                         goto open_again;
 727                 }
 728         }
 729
 730         return 0;
 731 }
 732
 733 /*
 734  * This function i.e. get_file_size() is the default .get_file_size
 735  * implementation of majority of I/O engines.
 736  */
 737 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 738 {
 739         return get_file_size(td, f);
 740 }
 741
 742 /*
 743  * open/close all files, so that ->real_file_size gets set
 744  */
 745 static int get_file_sizes(struct thread_data *td)
 746 {
 747         struct fio_file *f;
 748         unsigned int i;
 749         int err = 0;
 750
 751         for_each_file(td, f, i) {
 752                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 753                                                                 f->file_name);
 754
 755                 if (td_io_get_file_size(td, f)) {
 756                         if (td->error != ENOENT) {
 757                                 log_err("%s\n", td->verror);
 758                                 err = 1;
 759                                 break;
 760                         }
 761                         clear_error(td);
 762                 }
 763
 764                 /*
 765                  * There are corner cases where we end up with -1 for
 766                  * ->real_file_size due to unsupported file type, etc.
 767                  * We then just set to size option value divided by number
 768                  * of files, similar to the way file ->io_size is set.
 769                  * stat(2) failure doesn't set ->real_file_size to -1.
 770                  */
 771                 if (f->real_file_size == -1ULL && td->o.size)
 772                         f->real_file_size = td->o.size / td->o.nr_files;
 773         }
 774
 775         return err;
 776 }
 777
 778 struct fio_mount {
 779         struct flist_head list;
 780         const char *base;
 781         char __base[256];
 782         unsigned int key;
 783 };
 784
 785 /*
 786  * Get free number of bytes for each file on each unique mount.
 787  */
 788 static unsigned long long get_fs_free_counts(struct thread_data *td)
 789 {
 790         struct flist_head *n, *tmp;
 791         unsigned long long ret = 0;
 792         struct fio_mount *fm;
 793         FLIST_HEAD(list);
 794         struct fio_file *f;
 795         unsigned int i;
 796
 797         for_each_file(td, f, i) {
 798                 struct stat sb;
 799                 char buf[256];
 800
 801                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 802                         if (f->real_file_size != -1ULL)
 803                                 ret += f->real_file_size;
 804                         continue;
 805                 } else if (f->filetype != FIO_TYPE_FILE)
 806                         continue;
 807
 808                 buf[255] = '\0';
 809                 strncpy(buf, f->file_name, 255);
 810
 811                 if (stat(buf, &sb) < 0) {
 812                         if (errno != ENOENT)
 813                                 break;
 814                         strcpy(buf, ".");
 815                         if (stat(buf, &sb) < 0)
 816                                 break;
 817                 }
 818
 819                 fm = NULL;
 820                 flist_for_each(n, &list) {
 821                         fm = flist_entry(n, struct fio_mount, list);
 822                         if (fm->key == sb.st_dev)
 823                                 break;
 824
 825                         fm = NULL;
 826                 }
 827
 828                 if (fm)
 829                         continue;
 830
 831                 fm = calloc(1, sizeof(*fm));
 832                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 833                 fm->base = basename(fm->__base);
 834                 fm->key = sb.st_dev;
 835                 flist_add(&fm->list, &list);
 836         }
 837
 838         flist_for_each_safe(n, tmp, &list) {
 839                 unsigned long long sz;
 840
 841                 fm = flist_entry(n, struct fio_mount, list);
 842                 flist_del(&fm->list);
 843
 844                 sz = get_fs_free_size(fm->base);
 845                 if (sz && sz != -1ULL)
 846                         ret += sz;
 847
 848                 free(fm);
 849         }
 850
 851         return ret;
 852 }
 853
 854 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 855 {
 856         struct thread_options *o = &td->o;
 857         unsigned long long align_bs;
 858         unsigned long long offset;
 859
 860         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 861                 return f->real_file_size;
 862
 863         if (o->start_offset_percent > 0) {
 864                 /*
 865                  * if offset_align is provided, set initial offset
 866                  */
 867                 if (fio_option_is_set(o, start_offset_align)) {
 868                         align_bs = o->start_offset_align;
 869                 } else {
 870                         /* else take the minimum block size */
 871                         align_bs = td_min_bs(td);
 872                 }
 873
 874                 /* calculate the raw offset */
 875                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 876                         (td->subjob_number * o->offset_increment);
 877
 878                 /*
 879                  * block align the offset at the next available boundary at
 880                  * ceiling(offset / align_bs) * align_bs
 881                  */
 882                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 883
 884         } else {
 885                 /* start_offset_percent not set */
 886                 offset = o->start_offset +
 887                                 td->subjob_number * o->offset_increment;
 888         }
 889
 890         return offset;
 891 }
 892
 893 /*
 894  * Open the files and setup files sizes, creating files if necessary.
 895  */
 896 int setup_files(struct thread_data *td)
 897 {
 898         unsigned long long total_size, extend_size;
 899         struct thread_options *o = &td->o;
 900         struct fio_file *f;
 901         unsigned int i, nr_fs_extra = 0;
 902         int err = 0, need_extend;
 903         int old_state;
 904         const unsigned long long bs = td_min_bs(td);
 905         uint64_t fs = 0;
 906
 907         dprint(FD_FILE, "setup files\n");
 908
 909         old_state = td_bump_runstate(td, TD_SETTING_UP);
 910
 911         if (o->read_iolog_file)
 912                 goto done;
 913
 914         /*
 915          * Find out physical size of files or devices for this thread,
 916          * before we determine I/O size and range of our targets.
 917          * If ioengine defines a setup() method, it's responsible for
 918          * opening the files and setting f->real_file_size to indicate
 919          * the valid range for that file.
 920          */
 921         if (td->io_ops->setup)
 922                 err = td->io_ops->setup(td);
 923         else
 924                 err = get_file_sizes(td);
 925
 926         if (err)
 927                 goto err_out;
 928
 929         /*
 930          * check sizes. if the files/devices do not exist and the size
 931          * isn't passed to fio, abort.
 932          */
 933         total_size = 0;
 934         for_each_file(td, f, i) {
 935                 f->fileno = i;
 936                 if (f->real_file_size == -1ULL)
 937                         total_size = -1ULL;
 938                 else
 939                         total_size += f->real_file_size;
 940         }
 941
 942         if (o->fill_device)
 943                 td->fill_device_size = get_fs_free_counts(td);
 944
 945         /*
 946          * device/file sizes are zero and no size given, punt
 947          */
 948         if ((!total_size || total_size == -1ULL) && !o->size &&
 949             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 950             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 951                 log_err("%s: you need to specify size=\n", o->name);
 952                 td_verror(td, EINVAL, "total_file_size");
 953                 goto err_out;
 954         }
 955
 956         /*
 957          * Calculate per-file size and potential extra size for the
 958          * first files, if needed (i.e. if we don't have a fixed size).
 959          */
 960         if (!o->file_size_low && o->nr_files) {
 961                 uint64_t all_fs;
 962
 963                 fs = o->size / o->nr_files;
 964                 all_fs = fs * o->nr_files;
 965
 966                 if (all_fs < o->size)
 967                         nr_fs_extra = (o->size - all_fs) / bs;
 968         }
 969
 970         /*
 971          * now file sizes are known, so we can set ->io_size. if size= is
 972          * not given, ->io_size is just equal to ->real_file_size. if size
 973          * is given, ->io_size is size / nr_files.
 974          */
 975         extend_size = total_size = 0;
 976         need_extend = 0;
 977         for_each_file(td, f, i) {
 978                 f->file_offset = get_start_offset(td, f);
 979
 980                 /*
 981                  * Update ->io_size depending on options specified.
 982                  * ->file_size_low being 0 means filesize option isn't set.
 983                  * Non zero ->file_size_low equals ->file_size_high means
 984                  * filesize option is set in a fixed size format.
 985                  * Non zero ->file_size_low not equals ->file_size_high means
 986                  * filesize option is set in a range format.
 987                  */
 988                 if (!o->file_size_low) {
 989                         /*
 990                          * no file size or range given, file size is equal to
 991                          * total size divided by number of files. If the size
 992                          * doesn't divide nicely with the min blocksize,
 993                          * make the first files bigger.
 994                          */
 995                         f->io_size = fs;
 996                         if (nr_fs_extra) {
 997                                 nr_fs_extra--;
 998                                 f->io_size += bs;
 999                         }
1000
1001                         /*
1002                          * We normally don't come here for regular files, but
1003                          * if the result is 0 for a regular file, set it to the
1004                          * real file size. This could be size of the existing
1005                          * one if it already exists, but otherwise will be set
1006                          * to 0. A new file won't be created because
1007                          * ->io_size + ->file_offset equals ->real_file_size.
1008                          */
1009                         if (!f->io_size) {
1010                                 if (f->file_offset > f->real_file_size)
1011                                         goto err_offset;
1012                                 f->io_size = f->real_file_size - f->file_offset;
1013                                 if (!f->io_size)
1014                                         log_info("fio: file %s may be ignored\n",
1015                                                 f->file_name);
1016                         }
1017                 } else if (f->real_file_size < o->file_size_low ||
1018                            f->real_file_size > o->file_size_high) {
1019                         if (f->file_offset > o->file_size_low)
1020                                 goto err_offset;
1021                         /*
1022                          * file size given. if it's fixed, use that. if it's a
1023                          * range, generate a random size in-between.
1024                          */
1025                         if (o->file_size_low == o->file_size_high)
1026                                 f->io_size = o->file_size_low - f->file_offset;
1027                         else {
1028                                 f->io_size = get_rand_file_size(td)
1029                                                 - f->file_offset;
1030                         }
1031                 } else
1032                         f->io_size = f->real_file_size - f->file_offset;
1033
1034                 if (f->io_size == -1ULL)
1035                         total_size = -1ULL;
1036                 else {
1037                         if (o->size_percent && o->size_percent != 100) {
1038                                 uint64_t file_size;
1039
1040                                 file_size = f->io_size + f->file_offset;
1041                                 f->io_size = (file_size *
1042                                               o->size_percent) / 100;
1043                                 if (f->io_size > (file_size - f->file_offset))
1044                                         f->io_size = file_size - f->file_offset;
1045
1046                                 f->io_size -= (f->io_size % td_min_bs(td));
1047                         }
1048                         total_size += f->io_size;
1049                 }
1050
1051                 if (f->filetype == FIO_TYPE_FILE &&
1052                     (f->io_size + f->file_offset) > f->real_file_size &&
1053                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1054                         if (!o->create_on_open) {
1055                                 need_extend++;
1056                                 extend_size += (f->io_size + f->file_offset);
1057                                 fio_file_set_extend(f);
1058                         } else
1059                                 f->real_file_size = f->io_size + f->file_offset;
1060                 }
1061         }
1062
1063         if (td->o.block_error_hist) {
1064                 int len;
1065
1066                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1067                 f = td->files[0];
1068                 len = f->io_size / td->o.bs[DDIR_TRIM];
1069                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1070                         log_err("fio: cannot calculate block histogram with "
1071                                 "%d trim blocks, maximum %d\n",
1072                                 len, MAX_NR_BLOCK_INFOS);
1073                         td_verror(td, EINVAL, "block_error_hist");
1074                         goto err_out;
1075                 }
1076
1077                 td->ts.nr_block_infos = len;
1078                 for (i = 0; i < len; i++)
1079                         td->ts.block_infos[i] =
1080                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1081         } else
1082                 td->ts.nr_block_infos = 0;
1083
1084         if (!o->size || (total_size && o->size > total_size))
1085                 o->size = total_size;
1086
1087         if (o->size < td_min_bs(td)) {
1088                 log_err("fio: blocksize too large for data set\n");
1089                 goto err_out;
1090         }
1091
1092         /*
1093          * See if we need to extend some files, typically needed when our
1094          * target regular files don't exist yet, but our jobs require them
1095          * initially due to read I/Os.
1096          */
1097         if (need_extend) {
1098                 temp_stall_ts = 1;
1099                 if (output_format & FIO_OUTPUT_NORMAL) {
1100                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1101                                  o->name,
1102                                  need_extend > 1 ? "s" : "",
1103                                  need_extend,
1104                                  need_extend > 1 ? "s" : "",
1105                                  need_extend > 1 ? "total " : "",
1106                                  extend_size >> 20);
1107                 }
1108
1109                 for_each_file(td, f, i) {
1110                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1111
1112                         if (!fio_file_extend(f))
1113                                 continue;
1114
1115                         assert(f->filetype == FIO_TYPE_FILE);
1116                         fio_file_clear_extend(f);
1117                         if (!o->fill_device) {
1118                                 old_len = f->real_file_size;
1119                                 extend_len = f->io_size + f->file_offset -
1120                                                 old_len;
1121                         }
1122                         f->real_file_size = (f->io_size + f->file_offset);
1123                         err = extend_file(td, f);
1124                         if (err)
1125                                 break;
1126
1127                         err = __file_invalidate_cache(td, f, old_len,
1128                                                                 extend_len);
1129
1130                         /*
1131                          * Shut up static checker
1132                          */
1133                         if (f->fd != -1)
1134                                 close(f->fd);
1135
1136                         f->fd = -1;
1137                         if (err)
1138                                 break;
1139                 }
1140                 temp_stall_ts = 0;
1141         }
1142
1143         if (err)
1144                 goto err_out;
1145
1146         /*
1147          * iolog already set the total io size, if we read back
1148          * stored entries.
1149          */
1150         if (!o->read_iolog_file) {
1151                 if (o->io_size)
1152                         td->total_io_size = o->io_size * o->loops;
1153                 else
1154                         td->total_io_size = o->size * o->loops;
1155         }
1156
1157 done:
1158         if (o->create_only)
1159                 td->done = 1;
1160
1161         td_restore_runstate(td, old_state);
1162
1163         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1164                 err = zbd_init(td);
1165                 if (err)
1166                         goto err_out;
1167         }
1168         return 0;
1169
1170 err_offset:
1171         log_err("%s: you need to specify valid offset=\n", o->name);
1172 err_out:
1173         td_restore_runstate(td, old_state);
1174         return 1;
1175 }
1176
1177 bool pre_read_files(struct thread_data *td)
1178 {
1179         struct fio_file *f;
1180         unsigned int i;
1181
1182         dprint(FD_FILE, "pre_read files\n");
1183
1184         for_each_file(td, f, i) {
1185                 if (!pre_read_file(td, f))
1186                         return false;
1187         }
1188
1189         return true;
1190 }
1191
1192 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1193 {
1194         unsigned int range_size, seed;
1195         unsigned long nranges;
1196         uint64_t fsize;
1197
1198         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1199         fsize = min(f->real_file_size, f->io_size);
1200
1201         nranges = (fsize + range_size - 1) / range_size;
1202
1203         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1204         if (!td->o.rand_repeatable)
1205                 seed = td->rand_seeds[4];
1206
1207         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1208                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1209         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1210                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1211         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1212                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1213 }
1214
1215 static bool init_rand_distribution(struct thread_data *td)
1216 {
1217         struct fio_file *f;
1218         unsigned int i;
1219         int state;
1220
1221         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1222                 return false;
1223
1224         state = td_bump_runstate(td, TD_SETTING_UP);
1225
1226         for_each_file(td, f, i)
1227                 __init_rand_distribution(td, f);
1228
1229         td_restore_runstate(td, state);
1230         return true;
1231 }
1232
1233 /*
1234  * Check if the number of blocks exceeds the randomness capability of
1235  * the selected generator. Tausworthe is 32-bit, the others are fullly
1236  * 64-bit capable.
1237  */
1238 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1239                                  uint64_t blocks)
1240 {
1241         if (blocks <= FRAND32_MAX)
1242                 return 0;
1243         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1244                 return 0;
1245
1246         /*
1247          * If the user hasn't specified a random generator, switch
1248          * to tausworthe64 with informational warning. If the user did
1249          * specify one, just warn.
1250          */
1251         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1252                         f->file_name);
1253
1254         if (!fio_option_is_set(&td->o, random_generator)) {
1255                 log_info("fio: Switching to tausworthe64. Use the "
1256                          "random_generator= option to get rid of this "
1257                          "warning.\n");
1258                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1259                 return 0;
1260         }
1261
1262         /*
1263          * Just make this information to avoid breaking scripts.
1264          */
1265         log_info("fio: Use the random_generator= option to switch to lfsr or "
1266                          "tausworthe64.\n");
1267         return 0;
1268 }
1269
1270 bool init_random_map(struct thread_data *td)
1271 {
1272         unsigned long long blocks;
1273         struct fio_file *f;
1274         unsigned int i;
1275
1276         if (init_rand_distribution(td))
1277                 return true;
1278         if (!td_random(td))
1279                 return true;
1280
1281         for_each_file(td, f, i) {
1282                 uint64_t fsize = min(f->real_file_size, f->io_size);
1283
1284                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1285
1286                 if (check_rand_gen_limits(td, f, blocks))
1287                         return false;
1288
1289                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1290                         unsigned long seed;
1291
1292                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1293
1294                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1295                                 fio_file_set_lfsr(f);
1296                                 continue;
1297                         }
1298                 } else if (!td->o.norandommap) {
1299                         f->io_axmap = axmap_new(blocks);
1300                         if (f->io_axmap) {
1301                                 fio_file_set_axmap(f);
1302                                 continue;
1303                         }
1304                 } else if (td->o.norandommap)
1305                         continue;
1306
1307                 if (!td->o.softrandommap) {
1308                         log_err("fio: failed allocating random map. If running"
1309                                 " a large number of jobs, try the 'norandommap'"
1310                                 " option or set 'softrandommap'. Or give"
1311                                 " a larger --alloc-size to fio.\n");
1312                         return false;
1313                 }
1314
1315                 log_info("fio: file %s failed allocating random map. Running "
1316                          "job without.\n", f->file_name);
1317         }
1318
1319         return true;
1320 }
1321
1322 void close_files(struct thread_data *td)
1323 {
1324         struct fio_file *f;
1325         unsigned int i;
1326
1327         for_each_file(td, f, i) {
1328                 if (fio_file_open(f))
1329                         td_io_close_file(td, f);
1330         }
1331 }
1332
1333 void close_and_free_files(struct thread_data *td)
1334 {
1335         struct fio_file *f;
1336         unsigned int i;
1337         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1338
1339         dprint(FD_FILE, "close files\n");
1340
1341         for_each_file(td, f, i) {
1342                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1343                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1344                         td_io_unlink_file(td, f);
1345                 }
1346
1347                 if (fio_file_open(f))
1348                         td_io_close_file(td, f);
1349
1350                 remove_file_hash(f);
1351
1352                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1353                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1354                         td_io_unlink_file(td, f);
1355                 }
1356
1357                 zbd_free_zone_info(f);
1358
1359                 if (use_free)
1360                         free(f->file_name);
1361                 else
1362                         sfree(f->file_name);
1363                 f->file_name = NULL;
1364                 if (fio_file_axmap(f)) {
1365                         axmap_free(f->io_axmap);
1366                         f->io_axmap = NULL;
1367                 }
1368                 if (use_free)
1369                         free(f);
1370                 else
1371                         sfree(f);
1372         }
1373
1374         td->o.filename = NULL;
1375         free(td->files);
1376         free(td->file_locks);
1377         td->files_index = 0;
1378         td->files = NULL;
1379         td->file_locks = NULL;
1380         td->o.file_lock_mode = FILE_LOCK_NONE;
1381         td->o.nr_files = 0;
1382 }
1383
1384 static void get_file_type(struct fio_file *f)
1385 {
1386         struct stat sb;
1387
1388         if (!strcmp(f->file_name, "-"))
1389                 f->filetype = FIO_TYPE_PIPE;
1390         else
1391                 f->filetype = FIO_TYPE_FILE;
1392
1393 #ifdef WIN32
1394         /* \\.\ is the device namespace in Windows, where every file is
1395          * a block device */
1396         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1397                 f->filetype = FIO_TYPE_BLOCK;
1398 #endif
1399
1400         if (!stat(f->file_name, &sb)) {
1401                 if (S_ISBLK(sb.st_mode))
1402                         f->filetype = FIO_TYPE_BLOCK;
1403                 else if (S_ISCHR(sb.st_mode))
1404                         f->filetype = FIO_TYPE_CHAR;
1405                 else if (S_ISFIFO(sb.st_mode))
1406                         f->filetype = FIO_TYPE_PIPE;
1407         }
1408 }
1409
1410 static bool __is_already_allocated(const char *fname, bool set)
1411 {
1412         struct flist_head *entry;
1413         bool ret;
1414
1415         ret = file_bloom_exists(fname, set);
1416         if (!ret)
1417                 return ret;
1418
1419         flist_for_each(entry, &filename_list) {
1420                 struct file_name *fn;
1421
1422                 fn = flist_entry(entry, struct file_name, list);
1423
1424                 if (!strcmp(fn->filename, fname))
1425                         return true;
1426         }
1427
1428         return false;
1429 }
1430
1431 static bool is_already_allocated(const char *fname)
1432 {
1433         bool ret;
1434
1435         fio_file_hash_lock();
1436         ret = __is_already_allocated(fname, false);
1437         fio_file_hash_unlock();
1438
1439         return ret;
1440 }
1441
1442 static void set_already_allocated(const char *fname)
1443 {
1444         struct file_name *fn;
1445
1446         fn = malloc(sizeof(struct file_name));
1447         fn->filename = strdup(fname);
1448
1449         fio_file_hash_lock();
1450         if (!__is_already_allocated(fname, true)) {
1451                 flist_add_tail(&fn->list, &filename_list);
1452                 fn = NULL;
1453         }
1454         fio_file_hash_unlock();
1455
1456         if (fn) {
1457                 free(fn->filename);
1458                 free(fn);
1459         }
1460 }
1461
1462 static void free_already_allocated(void)
1463 {
1464         struct flist_head *entry, *tmp;
1465         struct file_name *fn;
1466
1467         if (flist_empty(&filename_list))
1468                 return;
1469
1470         fio_file_hash_lock();
1471         flist_for_each_safe(entry, tmp, &filename_list) {
1472                 fn = flist_entry(entry, struct file_name, list);
1473                 free(fn->filename);
1474                 flist_del(&fn->list);
1475                 free(fn);
1476         }
1477
1478         fio_file_hash_unlock();
1479 }
1480
1481 static struct fio_file *alloc_new_file(struct thread_data *td)
1482 {
1483         struct fio_file *f;
1484
1485         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1486                 f = calloc(1, sizeof(*f));
1487         else
1488                 f = scalloc(1, sizeof(*f));
1489         if (!f) {
1490                 assert(0);
1491                 return NULL;
1492         }
1493
1494         f->fd = -1;
1495         f->shadow_fd = -1;
1496         fio_file_reset(td, f);
1497         return f;
1498 }
1499
1500 bool exists_and_not_regfile(const char *filename)
1501 {
1502         struct stat sb;
1503
1504         if (lstat(filename, &sb) == -1)
1505                 return false;
1506
1507 #ifndef WIN32 /* NOT Windows */
1508         if (S_ISREG(sb.st_mode))
1509                 return false;
1510 #else
1511         /* \\.\ is the device namespace in Windows, where every file
1512          * is a device node */
1513         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1514                 return false;
1515 #endif
1516
1517         return true;
1518 }
1519
1520 static bool create_work_dirs(struct thread_data *td, const char *fname)
1521 {
1522         char path[PATH_MAX];
1523         char *start, *end;
1524
1525         if (td->o.directory) {
1526                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1527                          FIO_OS_PATH_SEPARATOR, fname);
1528                 start = strstr(path, fname);
1529         } else {
1530                 snprintf(path, PATH_MAX, "%s", fname);
1531                 start = path;
1532         }
1533
1534         end = start;
1535         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1536                 if (end == start)
1537                         break;
1538                 *end = '\0';
1539                 errno = 0;
1540 #ifdef CONFIG_HAVE_MKDIR_TWO
1541                 if (mkdir(path, 0600) && errno != EEXIST) {
1542 #else
1543                 if (mkdir(path) && errno != EEXIST) {
1544 #endif
1545                         log_err("fio: failed to create dir (%s): %d\n",
1546                                 start, errno);
1547                         return false;
1548                 }
1549                 *end = FIO_OS_PATH_SEPARATOR;
1550                 end++;
1551         }
1552         td->flags |= TD_F_DIRS_CREATED;
1553         return true;
1554 }
1555
1556 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1557 {
1558         int cur_files = td->files_index;
1559         char file_name[PATH_MAX];
1560         struct fio_file *f;
1561         int len = 0;
1562
1563         dprint(FD_FILE, "add file %s\n", fname);
1564
1565         if (td->o.directory)
1566                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1567                                         td->o.unique_filename);
1568
1569         sprintf(file_name + len, "%s", fname);
1570
1571         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1572             !(td->flags & TD_F_DIRS_CREATED) &&
1573             !create_work_dirs(td, fname))
1574                 return 1;
1575
1576         /* clean cloned siblings using existing files */
1577         if (numjob && is_already_allocated(file_name) &&
1578             !exists_and_not_regfile(fname))
1579                 return 0;
1580
1581         f = alloc_new_file(td);
1582
1583         if (td->files_size <= td->files_index) {
1584                 unsigned int new_size = td->o.nr_files + 1;
1585
1586                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1587
1588                 td->files = realloc(td->files, new_size * sizeof(f));
1589                 if (td->files == NULL) {
1590                         log_err("fio: realloc OOM\n");
1591                         assert(0);
1592                 }
1593                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1594                         td->file_locks = realloc(td->file_locks, new_size);
1595                         if (!td->file_locks) {
1596                                 log_err("fio: realloc OOM\n");
1597                                 assert(0);
1598                         }
1599                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1600                 }
1601                 td->files_size = new_size;
1602         }
1603         td->files[cur_files] = f;
1604         f->fileno = cur_files;
1605
1606         /*
1607          * init function, io engine may not be loaded yet
1608          */
1609         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1610                 f->real_file_size = -1ULL;
1611
1612         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1613                 f->file_name = strdup(file_name);
1614         else
1615                 f->file_name = smalloc_strdup(file_name);
1616
1617         /* can't handle smalloc failure from here */
1618         assert(f->file_name);
1619
1620         get_file_type(f);
1621
1622         switch (td->o.file_lock_mode) {
1623         case FILE_LOCK_NONE:
1624                 break;
1625         case FILE_LOCK_READWRITE:
1626                 f->rwlock = fio_rwlock_init();
1627                 break;
1628         case FILE_LOCK_EXCLUSIVE:
1629                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1630                 break;
1631         default:
1632                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1633                 assert(0);
1634         }
1635
1636         td->files_index++;
1637
1638         if (td->o.numjobs > 1)
1639                 set_already_allocated(file_name);
1640
1641         if (inc)
1642                 td->o.nr_files++;
1643
1644         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1645                                                         cur_files);
1646
1647         return cur_files;
1648 }
1649
1650 int add_file_exclusive(struct thread_data *td, const char *fname)
1651 {
1652         struct fio_file *f;
1653         unsigned int i;
1654
1655         for_each_file(td, f, i) {
1656                 if (!strcmp(f->file_name, fname))
1657                         return i;
1658         }
1659
1660         return add_file(td, fname, 0, 1);
1661 }
1662
1663 void get_file(struct fio_file *f)
1664 {
1665         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1666         assert(fio_file_open(f));
1667         f->references++;
1668 }
1669
1670 int put_file(struct thread_data *td, struct fio_file *f)
1671 {
1672         int f_ret = 0, ret = 0;
1673
1674         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1675
1676         if (!fio_file_open(f)) {
1677                 assert(f->fd == -1);
1678                 return 0;
1679         }
1680
1681         assert(f->references);
1682         if (--f->references)
1683                 return 0;
1684
1685         disk_util_dec(f->du);
1686
1687         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1688                 unlock_file_all(td, f);
1689
1690         if (should_fsync(td) && td->o.fsync_on_close) {
1691                 f_ret = fsync(f->fd);
1692                 if (f_ret < 0)
1693                         f_ret = errno;
1694         }
1695
1696         if (td->io_ops->close_file)
1697                 ret = td->io_ops->close_file(td, f);
1698
1699         if (!ret)
1700                 ret = f_ret;
1701
1702         td->nr_open_files--;
1703         fio_file_clear_closing(f);
1704         fio_file_clear_open(f);
1705         assert(f->fd == -1);
1706         return ret;
1707 }
1708
1709 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1710 {
1711         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1712                 return;
1713
1714         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1715                 if (ddir == DDIR_READ)
1716                         fio_rwlock_read(f->rwlock);
1717                 else
1718                         fio_rwlock_write(f->rwlock);
1719         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1720                 fio_sem_down(f->lock);
1721
1722         td->file_locks[f->fileno] = td->o.file_lock_mode;
1723 }
1724
1725 void unlock_file(struct thread_data *td, struct fio_file *f)
1726 {
1727         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1728                 return;
1729
1730         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1731                 fio_rwlock_unlock(f->rwlock);
1732         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1733                 fio_sem_up(f->lock);
1734
1735         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1736 }
1737
1738 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1739 {
1740         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1741                 return;
1742         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1743                 unlock_file(td, f);
1744 }
1745
1746 static bool recurse_dir(struct thread_data *td, const char *dirname)
1747 {
1748         struct dirent *dir;
1749         bool ret = false;
1750         DIR *D;
1751
1752         D = opendir(dirname);
1753         if (!D) {
1754                 char buf[FIO_VERROR_SIZE];
1755
1756                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1757                 td_verror(td, errno, buf);
1758                 return true;
1759         }
1760
1761         while ((dir = readdir(D)) != NULL) {
1762                 char full_path[PATH_MAX];
1763                 struct stat sb;
1764
1765                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1766                         continue;
1767
1768                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1769
1770                 if (lstat(full_path, &sb) == -1) {
1771                         if (errno != ENOENT) {
1772                                 td_verror(td, errno, "stat");
1773                                 ret = true;
1774                                 break;
1775                         }
1776                 }
1777
1778                 if (S_ISREG(sb.st_mode)) {
1779                         add_file(td, full_path, 0, 1);
1780                         continue;
1781                 }
1782                 if (!S_ISDIR(sb.st_mode))
1783                         continue;
1784
1785                 ret = recurse_dir(td, full_path);
1786                 if (ret)
1787                         break;
1788         }
1789
1790         closedir(D);
1791         return ret;
1792 }
1793
1794 int add_dir_files(struct thread_data *td, const char *path)
1795 {
1796         int ret = recurse_dir(td, path);
1797
1798         if (!ret)
1799                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1800
1801         return ret;
1802 }
1803
1804 void dup_files(struct thread_data *td, struct thread_data *org)
1805 {
1806         struct fio_file *f;
1807         unsigned int i;
1808
1809         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1810
1811         if (!org->files)
1812                 return;
1813
1814         td->files = malloc(org->files_index * sizeof(f));
1815
1816         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1817                 td->file_locks = malloc(org->files_index);
1818
1819         for_each_file(org, f, i) {
1820                 struct fio_file *__f;
1821
1822                 __f = alloc_new_file(td);
1823
1824                 if (f->file_name) {
1825                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1826                                 __f->file_name = strdup(f->file_name);
1827                         else
1828                                 __f->file_name = smalloc_strdup(f->file_name);
1829
1830                         /* can't handle smalloc failure from here */
1831                         assert(__f->file_name);
1832                         __f->filetype = f->filetype;
1833                 }
1834
1835                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1836                         __f->lock = f->lock;
1837                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1838                         __f->rwlock = f->rwlock;
1839
1840                 td->files[i] = __f;
1841         }
1842 }
1843
1844 /*
1845  * Returns the index that matches the filename, or -1 if not there
1846  */
1847 int get_fileno(struct thread_data *td, const char *fname)
1848 {
1849         struct fio_file *f;
1850         unsigned int i;
1851
1852         for_each_file(td, f, i)
1853                 if (!strcmp(f->file_name, fname))
1854                         return i;
1855
1856         return -1;
1857 }
1858
1859 /*
1860  * For log usage, where we add/open/close files automatically
1861  */
1862 void free_release_files(struct thread_data *td)
1863 {
1864         close_files(td);
1865         td->o.nr_files = 0;
1866         td->o.open_files = 0;
1867         td->files_index = 0;
1868 }
1869
1870 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1871 {
1872         int i;
1873
1874         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1875                 f->last_pos[i] = f->file_offset;
1876                 f->last_start[i] = -1ULL;
1877         }
1878
1879         if (fio_file_axmap(f))
1880                 axmap_reset(f->io_axmap);
1881         else if (fio_file_lfsr(f))
1882                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1883
1884         zbd_file_reset(td, f);
1885 }
1886
1887 bool fio_files_done(struct thread_data *td)
1888 {
1889         struct fio_file *f;
1890         unsigned int i;
1891
1892         for_each_file(td, f, i)
1893                 if (!fio_file_done(f))
1894                         return false;
1895
1896         return true;
1897 }
1898
1899 /* free memory used in initialization phase only */
1900 void filesetup_mem_free(void)
1901 {
1902         free_already_allocated();
1903 }
1904
1905 /*
1906  * This function is for platforms which support direct I/O but not O_DIRECT.
1907  */
1908 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1909 {
1910 #ifdef FIO_OS_DIRECTIO
1911         int ret = fio_set_odirect(f);
1912
1913         if (ret) {
1914                 td_verror(td, ret, "fio_set_directio");
1915 #if defined(__sun__)
1916                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1917                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1918                 } else {
1919                         log_err("fio: the file system does not seem to support direct IO\n");
1920                 }
1921 #else
1922                 log_err("fio: the file system does not seem to support direct IO\n");
1923 #endif
1924                 return -1;
1925         }
1926
1927         return 0;
1928 #else
1929         log_err("fio: direct IO is not supported on this host operating system\n");
1930         return -1;
1931 #endif
1932 }