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         case FIO_FALLOCATE_TRUNCATE: {
  99                 int r;
 100
 101                 dprint(FD_FILE, "ftruncate file %s size %llu\n",
 102                                 f->file_name,
 103                                 (unsigned long long) f->real_file_size);
 104                 r = ftruncate(f->fd, f->real_file_size);
 105                 if (r != 0)
 106                         td_verror(td, errno, "ftruncate");
 107
 108                 break;
 109         }
 110         default:
 111                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
 112                 assert(0);
 113         }
 114 }
 115
 116 /*
 117  * Leaves f->fd open on success, caller must close
 118  */
 119 static int extend_file(struct thread_data *td, struct fio_file *f)
 120 {
 121         int new_layout = 0, unlink_file = 0, flags;
 122         unsigned long long left;
 123         unsigned long long bs;
 124         char *b = NULL;
 125
 126         if (read_only) {
 127                 log_err("fio: refusing extend of file due to read-only\n");
 128                 return 0;
 129         }
 130
 131         /*
 132          * check if we need to lay the file out complete again. fio
 133          * does that for operations involving reads, or for writes
 134          * where overwrite is set
 135          */
 136         if (td_read(td) ||
 137            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
 138             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
 139                 new_layout = 1;
 140         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
 141                 unlink_file = 1;
 142
 143         if (unlink_file || new_layout) {
 144                 int ret;
 145
 146                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
 147
 148                 ret = td_io_unlink_file(td, f);
 149                 if (ret != 0 && ret != ENOENT) {
 150                         td_verror(td, errno, "unlink");
 151                         return 1;
 152                 }
 153         }
 154
 155         flags = O_WRONLY;
 156         if (td->o.allow_create)
 157                 flags |= O_CREAT;
 158         if (new_layout)
 159                 flags |= O_TRUNC;
 160
 161 #ifdef WIN32
 162         flags |= _O_BINARY;
 163 #endif
 164
 165         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 166         f->fd = open(f->file_name, flags, 0644);
 167         if (f->fd < 0) {
 168                 int err = errno;
 169
 170                 if (err == ENOENT && !td->o.allow_create)
 171                         log_err("fio: file creation disallowed by "
 172                                         "allow_file_create=0\n");
 173                 else
 174                         td_verror(td, err, "open");
 175                 return 1;
 176         }
 177
 178         fallocate_file(td, f);
 179
 180         /*
 181          * If our jobs don't require regular files initially, we're done.
 182          */
 183         if (!new_layout)
 184                 goto done;
 185
 186         /*
 187          * The size will be -1ULL when fill_device is used, so don't truncate
 188          * or fallocate this file, just write it
 189          */
 190         if (!td->o.fill_device) {
 191                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 192                                         (unsigned long long) f->real_file_size);
 193                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 194                         if (errno != EFBIG) {
 195                                 td_verror(td, errno, "ftruncate");
 196                                 goto err;
 197                         }
 198                 }
 199         }
 200
 201         left = f->real_file_size;
 202         bs = td->o.max_bs[DDIR_WRITE];
 203         if (bs > left)
 204                 bs = left;
 205
 206         b = malloc(bs);
 207         if (!b) {
 208                 td_verror(td, errno, "malloc");
 209                 goto err;
 210         }
 211
 212         while (left && !td->terminate) {
 213                 ssize_t r;
 214
 215                 if (bs > left)
 216                         bs = left;
 217
 218                 fill_io_buffer(td, b, bs, bs);
 219
 220                 r = write(f->fd, b, bs);
 221
 222                 if (r > 0) {
 223                         left -= r;
 224                         continue;
 225                 } else {
 226                         if (r < 0) {
 227                                 int __e = errno;
 228
 229                                 if (__e == ENOSPC) {
 230                                         if (td->o.fill_device)
 231                                                 break;
 232                                         log_info("fio: ENOSPC on laying out "
 233                                                  "file, stopping\n");
 234                                         break;
 235                                 }
 236                                 td_verror(td, errno, "write");
 237                         } else
 238                                 td_verror(td, EIO, "write");
 239
 240                         break;
 241                 }
 242         }
 243
 244         if (td->terminate) {
 245                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 246                 td_io_unlink_file(td, f);
 247         } else if (td->o.create_fsync) {
 248                 if (fsync(f->fd) < 0) {
 249                         td_verror(td, errno, "fsync");
 250                         goto err;
 251                 }
 252         }
 253         if (td->o.fill_device && !td_write(td)) {
 254                 fio_file_clear_size_known(f);
 255                 if (td_io_get_file_size(td, f))
 256                         goto err;
 257                 if (f->io_size > f->real_file_size)
 258                         f->io_size = f->real_file_size;
 259         }
 260
 261         free(b);
 262 done:
 263         return 0;
 264 err:
 265         close(f->fd);
 266         f->fd = -1;
 267         if (b)
 268                 free(b);
 269         return 1;
 270 }
 271
 272 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 273 {
 274         int r, did_open = 0, old_runstate;
 275         unsigned long long left;
 276         unsigned long long bs;
 277         bool ret = true;
 278         char *b;
 279
 280         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 281             td_ioengine_flagged(td, FIO_NOIO))
 282                 return true;
 283
 284         if (f->filetype == FIO_TYPE_CHAR)
 285                 return true;
 286
 287         if (!fio_file_open(f)) {
 288                 if (td->io_ops->open_file(td, f)) {
 289                         log_err("fio: cannot pre-read, failed to open file\n");
 290                         return false;
 291                 }
 292                 did_open = 1;
 293         }
 294
 295         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 296
 297         left = f->io_size;
 298         bs = td->o.max_bs[DDIR_READ];
 299         if (bs > left)
 300                 bs = left;
 301
 302         b = malloc(bs);
 303         if (!b) {
 304                 td_verror(td, errno, "malloc");
 305                 ret = false;
 306                 goto error;
 307         }
 308         memset(b, 0, bs);
 309
 310         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 311                 td_verror(td, errno, "lseek");
 312                 log_err("fio: failed to lseek pre-read file\n");
 313                 ret = false;
 314                 goto error;
 315         }
 316
 317         while (left && !td->terminate) {
 318                 if (bs > left)
 319                         bs = left;
 320
 321                 r = read(f->fd, b, bs);
 322
 323                 if (r == (int) bs) {
 324                         left -= bs;
 325                         continue;
 326                 } else {
 327                         td_verror(td, EIO, "pre_read");
 328                         break;
 329                 }
 330         }
 331
 332 error:
 333         td_restore_runstate(td, old_runstate);
 334
 335         if (did_open)
 336                 td->io_ops->close_file(td, f);
 337
 338         free(b);
 339         return ret;
 340 }
 341
 342 unsigned long long get_rand_file_size(struct thread_data *td)
 343 {
 344         unsigned long long ret, sized;
 345         uint64_t frand_max;
 346         uint64_t r;
 347
 348         frand_max = rand_max(&td->file_size_state);
 349         r = __rand(&td->file_size_state);
 350         sized = td->o.file_size_high - td->o.file_size_low;
 351         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 352         ret += td->o.file_size_low;
 353         ret -= (ret % td->o.rw_min_bs);
 354         return ret;
 355 }
 356
 357 static int file_size(struct thread_data *td, struct fio_file *f)
 358 {
 359         struct stat st;
 360
 361         if (stat(f->file_name, &st) == -1) {
 362                 td_verror(td, errno, "fstat");
 363                 return 1;
 364         }
 365
 366         f->real_file_size = st.st_size;
 367         return 0;
 368 }
 369
 370 static int bdev_size(struct thread_data *td, struct fio_file *f)
 371 {
 372         unsigned long long bytes = 0;
 373         int r;
 374
 375         if (td->io_ops->open_file(td, f)) {
 376                 log_err("fio: failed opening blockdev %s for size check\n",
 377                         f->file_name);
 378                 return 1;
 379         }
 380
 381         r = blockdev_size(f, &bytes);
 382         if (r) {
 383                 td_verror(td, r, "blockdev_size");
 384                 goto err;
 385         }
 386
 387         if (!bytes) {
 388                 log_err("%s: zero sized block device?\n", f->file_name);
 389                 goto err;
 390         }
 391
 392         f->real_file_size = bytes;
 393         td->io_ops->close_file(td, f);
 394         return 0;
 395 err:
 396         td->io_ops->close_file(td, f);
 397         return 1;
 398 }
 399
 400 static int char_size(struct thread_data *td, struct fio_file *f)
 401 {
 402 #ifdef FIO_HAVE_CHARDEV_SIZE
 403         unsigned long long bytes = 0;
 404         int r;
 405
 406         if (td->io_ops->open_file(td, f)) {
 407                 log_err("fio: failed opening chardev %s for size check\n",
 408                         f->file_name);
 409                 return 1;
 410         }
 411
 412         r = chardev_size(f, &bytes);
 413         if (r) {
 414                 td_verror(td, r, "chardev_size");
 415                 goto err;
 416         }
 417
 418         if (!bytes) {
 419                 log_err("%s: zero sized char device?\n", f->file_name);
 420                 goto err;
 421         }
 422
 423         f->real_file_size = bytes;
 424         td->io_ops->close_file(td, f);
 425         return 0;
 426 err:
 427         td->io_ops->close_file(td, f);
 428         return 1;
 429 #else
 430         f->real_file_size = -1ULL;
 431         return 0;
 432 #endif
 433 }
 434
 435 static int get_file_size(struct thread_data *td, struct fio_file *f)
 436 {
 437         int ret = 0;
 438
 439         if (fio_file_size_known(f))
 440                 return 0;
 441
 442         if (f->filetype == FIO_TYPE_FILE)
 443                 ret = file_size(td, f);
 444         else if (f->filetype == FIO_TYPE_BLOCK)
 445                 ret = bdev_size(td, f);
 446         else if (f->filetype == FIO_TYPE_CHAR)
 447                 ret = char_size(td, f);
 448         else {
 449                 f->real_file_size = -1;
 450                 log_info("%s: failed to get file size of %s\n", td->o.name,
 451                                         f->file_name);
 452                 return 1; /* avoid offset extends end error message */
 453         }
 454
 455         /*
 456          * Leave ->real_file_size with 0 since it could be expectation
 457          * of initial setup for regular files.
 458          */
 459         if (ret)
 460                 return ret;
 461
 462         /*
 463          * ->file_offset normally hasn't been initialized yet, so this
 464          * is basically always false unless ->real_file_size is -1, but
 465          * if ->real_file_size is -1 this message doesn't make sense.
 466          * As a result, this message is basically useless.
 467          */
 468         if (f->file_offset > f->real_file_size) {
 469                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 470                                         (unsigned long long) f->file_offset,
 471                                         (unsigned long long) f->real_file_size);
 472                 return 1;
 473         }
 474
 475         fio_file_set_size_known(f);
 476         return 0;
 477 }
 478
 479 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 480                                    unsigned long long off,
 481                                    unsigned long long len)
 482 {
 483         int errval = 0, ret = 0;
 484
 485 #ifdef CONFIG_ESX
 486         return 0;
 487 #endif
 488
 489         if (len == -1ULL)
 490                 len = f->io_size;
 491         if (off == -1ULL)
 492                 off = f->file_offset;
 493
 494         if (len == -1ULL || off == -1ULL)
 495                 return 0;
 496
 497         if (td->io_ops->invalidate) {
 498                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 499                         f->file_name);
 500                 ret = td->io_ops->invalidate(td, f);
 501                 if (ret < 0)
 502                         errval = -ret;
 503         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 504                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
 505                        td->io_ops->name);
 506         } else if (f->filetype == FIO_TYPE_FILE) {
 507                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 508                         f->file_name, off, len);
 509                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 510                 if (ret)
 511                         errval = ret;
 512         } else if (f->filetype == FIO_TYPE_BLOCK) {
 513                 int retry_count = 0;
 514
 515                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 516                 ret = blockdev_invalidate_cache(f);
 517                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 518                         /*
 519                          * Linux multipath devices reject ioctl while
 520                          * the maps are being updated. That window can
 521                          * last tens of milliseconds; we'll try up to
 522                          * a quarter of a second.
 523                          */
 524                         usleep(10000);
 525                         ret = blockdev_invalidate_cache(f);
 526                 }
 527                 if (ret < 0 && errno == EACCES && geteuid()) {
 528                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 529                                 log_err("fio: only root may flush block "
 530                                         "devices. Cache flush bypassed!\n");
 531                         }
 532                 }
 533                 if (ret < 0)
 534                         errval = errno;
 535         } else if (f->filetype == FIO_TYPE_CHAR ||
 536                    f->filetype == FIO_TYPE_PIPE) {
 537                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 538         }
 539
 540         /*
 541          * Cache flushing isn't a fatal condition, and we know it will
 542          * happen on some platforms where we don't have the proper
 543          * function to flush eg block device caches. So just warn and
 544          * continue on our way.
 545          */
 546         if (errval)
 547                 log_info("fio: cache invalidation of %s failed: %s\n",
 548                          f->file_name, strerror(errval));
 549
 550         return 0;
 551
 552 }
 553
 554 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 555 {
 556         if (!fio_file_open(f))
 557                 return 0;
 558
 559         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 560 }
 561
 562 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 563 {
 564         int ret = 0;
 565
 566         dprint(FD_FILE, "fd close %s\n", f->file_name);
 567
 568         remove_file_hash(f);
 569
 570         if (close(f->fd) < 0)
 571                 ret = errno;
 572
 573         f->fd = -1;
 574
 575         if (f->shadow_fd != -1) {
 576                 close(f->shadow_fd);
 577                 f->shadow_fd = -1;
 578         }
 579
 580         f->engine_pos = 0;
 581         return ret;
 582 }
 583
 584 int file_lookup_open(struct fio_file *f, int flags)
 585 {
 586         struct fio_file *__f;
 587         int from_hash;
 588
 589         __f = lookup_file_hash(f->file_name);
 590         if (__f) {
 591                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 592                 f->lock = __f->lock;
 593                 from_hash = 1;
 594         } else {
 595                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 596                 from_hash = 0;
 597         }
 598
 599 #ifdef WIN32
 600         flags |= _O_BINARY;
 601 #endif
 602
 603         f->fd = open(f->file_name, flags, 0600);
 604         return from_hash;
 605 }
 606
 607 static int file_close_shadow_fds(struct thread_data *td)
 608 {
 609         struct fio_file *f;
 610         int num_closed = 0;
 611         unsigned int i;
 612
 613         for_each_file(td, f, i) {
 614                 if (f->shadow_fd == -1)
 615                         continue;
 616
 617                 close(f->shadow_fd);
 618                 f->shadow_fd = -1;
 619                 num_closed++;
 620         }
 621
 622         return num_closed;
 623 }
 624
 625 int generic_open_file(struct thread_data *td, struct fio_file *f)
 626 {
 627         int is_std = 0;
 628         int flags = 0;
 629         int from_hash = 0;
 630
 631         dprint(FD_FILE, "fd open %s\n", f->file_name);
 632
 633         if (!strcmp(f->file_name, "-")) {
 634                 if (td_rw(td)) {
 635                         log_err("fio: can't read/write to stdin/out\n");
 636                         return 1;
 637                 }
 638                 is_std = 1;
 639
 640                 /*
 641                  * move output logging to stderr, if we are writing to stdout
 642                  */
 643                 if (td_write(td))
 644                         f_out = stderr;
 645         }
 646
 647         if (td_trim(td))
 648                 goto skip_flags;
 649         if (td->o.odirect)
 650                 flags |= OS_O_DIRECT;
 651         if (td->o.oatomic) {
 652                 if (!FIO_O_ATOMIC) {
 653                         td_verror(td, EINVAL, "OS does not support atomic IO");
 654                         return 1;
 655                 }
 656                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 657         }
 658         if (td->o.sync_io)
 659                 flags |= O_SYNC;
 660         if (td->o.create_on_open && td->o.allow_create)
 661                 flags |= O_CREAT;
 662 skip_flags:
 663         if (f->filetype != FIO_TYPE_FILE)
 664                 flags |= FIO_O_NOATIME;
 665
 666 open_again:
 667         if (td_write(td)) {
 668                 if (!read_only)
 669                         flags |= O_RDWR;
 670
 671                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 672                         flags |= O_CREAT;
 673
 674                 if (is_std)
 675                         f->fd = dup(STDOUT_FILENO);
 676                 else
 677                         from_hash = file_lookup_open(f, flags);
 678         } else if (td_read(td)) {
 679                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 680                         flags |= O_RDWR;
 681                 else
 682                         flags |= O_RDONLY;
 683
 684                 if (is_std)
 685                         f->fd = dup(STDIN_FILENO);
 686                 else
 687                         from_hash = file_lookup_open(f, flags);
 688         } else if (td_trim(td)) {
 689                 assert(!td_rw(td)); /* should have matched above */
 690                 if (!read_only)
 691                         flags |= O_RDWR;
 692                 from_hash = file_lookup_open(f, flags);
 693         }
 694
 695         if (f->fd == -1) {
 696                 char buf[FIO_VERROR_SIZE];
 697                 int __e = errno;
 698
 699                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 700                         flags &= ~FIO_O_NOATIME;
 701                         goto open_again;
 702                 }
 703                 if (__e == EMFILE && file_close_shadow_fds(td))
 704                         goto open_again;
 705
 706                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 707
 708                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 709                         log_err("fio: looks like your file system does not " \
 710                                 "support direct=1/buffered=0\n");
 711                 }
 712
 713                 td_verror(td, __e, buf);
 714                 return 1;
 715         }
 716
 717         if (!from_hash && f->fd != -1) {
 718                 if (add_file_hash(f)) {
 719                         int fio_unused ret;
 720
 721                         /*
 722                          * Stash away descriptor for later close. This is to
 723                          * work-around a "feature" on Linux, where a close of
 724                          * an fd that has been opened for write will trigger
 725                          * udev to call blkid to check partitions, fs id, etc.
 726                          * That pollutes the device cache, which can slow down
 727                          * unbuffered accesses.
 728                          */
 729                         if (f->shadow_fd == -1)
 730                                 f->shadow_fd = f->fd;
 731                         else {
 732                                 /*
 733                                  * OK to ignore, we haven't done anything
 734                                  * with it
 735                                  */
 736                                 ret = generic_close_file(td, f);
 737                         }
 738                         goto open_again;
 739                 }
 740         }
 741
 742         return 0;
 743 }
 744
 745 /*
 746  * This function i.e. get_file_size() is the default .get_file_size
 747  * implementation of majority of I/O engines.
 748  */
 749 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 750 {
 751         return get_file_size(td, f);
 752 }
 753
 754 /*
 755  * open/close all files, so that ->real_file_size gets set
 756  */
 757 static int get_file_sizes(struct thread_data *td)
 758 {
 759         struct fio_file *f;
 760         unsigned int i;
 761         int err = 0;
 762
 763         for_each_file(td, f, i) {
 764                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 765                                                                 f->file_name);
 766
 767                 if (td_io_get_file_size(td, f)) {
 768                         if (td->error != ENOENT) {
 769                                 log_err("%s\n", td->verror);
 770                                 err = 1;
 771                                 break;
 772                         }
 773                         clear_error(td);
 774                 }
 775
 776                 /*
 777                  * There are corner cases where we end up with -1 for
 778                  * ->real_file_size due to unsupported file type, etc.
 779                  * We then just set to size option value divided by number
 780                  * of files, similar to the way file ->io_size is set.
 781                  * stat(2) failure doesn't set ->real_file_size to -1.
 782                  */
 783                 if (f->real_file_size == -1ULL && td->o.size)
 784                         f->real_file_size = td->o.size / td->o.nr_files;
 785         }
 786
 787         return err;
 788 }
 789
 790 struct fio_mount {
 791         struct flist_head list;
 792         const char *base;
 793         char __base[256];
 794         unsigned int key;
 795 };
 796
 797 /*
 798  * Get free number of bytes for each file on each unique mount.
 799  */
 800 static unsigned long long get_fs_free_counts(struct thread_data *td)
 801 {
 802         struct flist_head *n, *tmp;
 803         unsigned long long ret = 0;
 804         struct fio_mount *fm;
 805         FLIST_HEAD(list);
 806         struct fio_file *f;
 807         unsigned int i;
 808
 809         for_each_file(td, f, i) {
 810                 struct stat sb;
 811                 char buf[256];
 812
 813                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 814                         if (f->real_file_size != -1ULL)
 815                                 ret += f->real_file_size;
 816                         continue;
 817                 } else if (f->filetype != FIO_TYPE_FILE)
 818                         continue;
 819
 820                 snprintf(buf, ARRAY_SIZE(buf), "%s", f->file_name);
 821
 822                 if (stat(buf, &sb) < 0) {
 823                         if (errno != ENOENT)
 824                                 break;
 825                         strcpy(buf, ".");
 826                         if (stat(buf, &sb) < 0)
 827                                 break;
 828                 }
 829
 830                 fm = NULL;
 831                 flist_for_each(n, &list) {
 832                         fm = flist_entry(n, struct fio_mount, list);
 833                         if (fm->key == sb.st_dev)
 834                                 break;
 835
 836                         fm = NULL;
 837                 }
 838
 839                 if (fm)
 840                         continue;
 841
 842                 fm = calloc(1, sizeof(*fm));
 843                 snprintf(fm->__base, ARRAY_SIZE(fm->__base), "%s", buf);
 844                 fm->base = basename(fm->__base);
 845                 fm->key = sb.st_dev;
 846                 flist_add(&fm->list, &list);
 847         }
 848
 849         flist_for_each_safe(n, tmp, &list) {
 850                 unsigned long long sz;
 851
 852                 fm = flist_entry(n, struct fio_mount, list);
 853                 flist_del(&fm->list);
 854
 855                 sz = get_fs_free_size(fm->base);
 856                 if (sz && sz != -1ULL)
 857                         ret += sz;
 858
 859                 free(fm);
 860         }
 861
 862         return ret;
 863 }
 864
 865 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 866 {
 867         bool align = false;
 868         struct thread_options *o = &td->o;
 869         unsigned long long align_bs;
 870         unsigned long long offset;
 871         unsigned long long increment;
 872
 873         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 874                 return f->real_file_size;
 875
 876         if (o->offset_increment_percent) {
 877                 assert(!o->offset_increment);
 878                 increment = o->offset_increment_percent * f->real_file_size / 100;
 879                 align = true;
 880         } else
 881                 increment = o->offset_increment;
 882
 883         if (o->start_offset_percent > 0) {
 884                 /* calculate the raw offset */
 885                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 886                         (td->subjob_number * increment);
 887
 888                 align = true;
 889         } else {
 890                 /* start_offset_percent not set */
 891                 offset = o->start_offset +
 892                                 td->subjob_number * increment;
 893         }
 894
 895         if (align) {
 896                 /*
 897                  * if offset_align is provided, use it
 898                  */
 899                 if (fio_option_is_set(o, start_offset_align)) {
 900                         align_bs = o->start_offset_align;
 901                 } else {
 902                         /* else take the minimum block size */
 903                         align_bs = td_min_bs(td);
 904                 }
 905
 906                 /*
 907                  * block align the offset at the next available boundary at
 908                  * ceiling(offset / align_bs) * align_bs
 909                  */
 910                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 911         }
 912
 913         return offset;
 914 }
 915
 916 static bool create_work_dirs(struct thread_data *td, const char *fname)
 917 {
 918         char path[PATH_MAX];
 919         char *start, *end;
 920
 921         if (td->o.directory) {
 922                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
 923                          FIO_OS_PATH_SEPARATOR, fname);
 924                 start = strstr(path, fname);
 925         } else {
 926                 snprintf(path, PATH_MAX, "%s", fname);
 927                 start = path;
 928         }
 929
 930         end = start;
 931         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
 932                 if (end == start)
 933                         break;
 934                 *end = '\0';
 935                 errno = 0;
 936 #ifdef CONFIG_HAVE_MKDIR_TWO
 937                 if (mkdir(path, 0600) && errno != EEXIST) {
 938 #else
 939                 if (mkdir(path) && errno != EEXIST) {
 940 #endif
 941                         log_err("fio: failed to create dir (%s): %d\n",
 942                                 start, errno);
 943                         return false;
 944                 }
 945                 *end = FIO_OS_PATH_SEPARATOR;
 946                 end++;
 947         }
 948         td->flags |= TD_F_DIRS_CREATED;
 949         return true;
 950 }
 951
 952 /*
 953  * Open the files and setup files sizes, creating files if necessary.
 954  */
 955 int setup_files(struct thread_data *td)
 956 {
 957         unsigned long long total_size, extend_size;
 958         struct thread_options *o = &td->o;
 959         struct fio_file *f;
 960         unsigned int i, nr_fs_extra = 0;
 961         int err = 0, need_extend;
 962         int old_state;
 963         const unsigned long long bs = td_min_bs(td);
 964         uint64_t fs = 0;
 965
 966         dprint(FD_FILE, "setup files\n");
 967
 968         old_state = td_bump_runstate(td, TD_SETTING_UP);
 969
 970         for_each_file(td, f, i) {
 971                 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
 972                     strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
 973                     !(td->flags & TD_F_DIRS_CREATED) &&
 974                     !create_work_dirs(td, f->file_name))
 975                         goto err_out;
 976         }
 977
 978         /*
 979          * Find out physical size of files or devices for this thread,
 980          * before we determine I/O size and range of our targets.
 981          * If ioengine defines a setup() method, it's responsible for
 982          * opening the files and setting f->real_file_size to indicate
 983          * the valid range for that file.
 984          */
 985         if (td->io_ops->setup)
 986                 err = td->io_ops->setup(td);
 987         else
 988                 err = get_file_sizes(td);
 989
 990         if (err)
 991                 goto err_out;
 992
 993         if (o->read_iolog_file)
 994                 goto done;
 995
 996         /*
 997          * check sizes. if the files/devices do not exist and the size
 998          * isn't passed to fio, abort.
 999          */
1000         total_size = 0;
1001         for_each_file(td, f, i) {
1002                 f->fileno = i;
1003                 if (f->real_file_size == -1ULL)
1004                         total_size = -1ULL;
1005                 else
1006                         total_size += f->real_file_size;
1007         }
1008
1009         if (o->fill_device)
1010                 td->fill_device_size = get_fs_free_counts(td);
1011
1012         /*
1013          * device/file sizes are zero and no size given, punt
1014          */
1015         if ((!total_size || total_size == -1ULL) && !o->size &&
1016             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1017             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1018                 log_err("%s: you need to specify size=\n", o->name);
1019                 td_verror(td, EINVAL, "total_file_size");
1020                 goto err_out;
1021         }
1022
1023         /*
1024          * Calculate per-file size and potential extra size for the
1025          * first files, if needed (i.e. if we don't have a fixed size).
1026          */
1027         if (!o->file_size_low && o->nr_files) {
1028                 uint64_t all_fs;
1029
1030                 fs = o->size / o->nr_files;
1031                 all_fs = fs * o->nr_files;
1032
1033                 if (all_fs < o->size)
1034                         nr_fs_extra = (o->size - all_fs) / bs;
1035         }
1036
1037         /*
1038          * now file sizes are known, so we can set ->io_size. if size= is
1039          * not given, ->io_size is just equal to ->real_file_size. if size
1040          * is given, ->io_size is size / nr_files.
1041          */
1042         extend_size = total_size = 0;
1043         need_extend = 0;
1044         for_each_file(td, f, i) {
1045                 f->file_offset = get_start_offset(td, f);
1046
1047                 /*
1048                  * Update ->io_size depending on options specified.
1049                  * ->file_size_low being 0 means filesize option isn't set.
1050                  * Non zero ->file_size_low equals ->file_size_high means
1051                  * filesize option is set in a fixed size format.
1052                  * Non zero ->file_size_low not equals ->file_size_high means
1053                  * filesize option is set in a range format.
1054                  */
1055                 if (!o->file_size_low) {
1056                         /*
1057                          * no file size or range given, file size is equal to
1058                          * total size divided by number of files. If the size
1059                          * doesn't divide nicely with the min blocksize,
1060                          * make the first files bigger.
1061                          */
1062                         f->io_size = fs;
1063                         if (nr_fs_extra) {
1064                                 nr_fs_extra--;
1065                                 f->io_size += bs;
1066                         }
1067
1068                         /*
1069                          * We normally don't come here for regular files, but
1070                          * if the result is 0 for a regular file, set it to the
1071                          * real file size. This could be size of the existing
1072                          * one if it already exists, but otherwise will be set
1073                          * to 0. A new file won't be created because
1074                          * ->io_size + ->file_offset equals ->real_file_size.
1075                          */
1076                         if (!f->io_size) {
1077                                 if (f->file_offset > f->real_file_size)
1078                                         goto err_offset;
1079                                 f->io_size = f->real_file_size - f->file_offset;
1080                                 if (!f->io_size)
1081                                         log_info("fio: file %s may be ignored\n",
1082                                                 f->file_name);
1083                         }
1084                 } else if (f->real_file_size < o->file_size_low ||
1085                            f->real_file_size > o->file_size_high) {
1086                         if (f->file_offset > o->file_size_low)
1087                                 goto err_offset;
1088                         /*
1089                          * file size given. if it's fixed, use that. if it's a
1090                          * range, generate a random size in-between.
1091                          */
1092                         if (o->file_size_low == o->file_size_high)
1093                                 f->io_size = o->file_size_low - f->file_offset;
1094                         else {
1095                                 f->io_size = get_rand_file_size(td)
1096                                                 - f->file_offset;
1097                         }
1098                 } else
1099                         f->io_size = f->real_file_size - f->file_offset;
1100
1101                 if (f->io_size == -1ULL)
1102                         total_size = -1ULL;
1103                 else {
1104                         if (o->size_percent && o->size_percent != 100) {
1105                                 uint64_t file_size;
1106
1107                                 file_size = f->io_size + f->file_offset;
1108                                 f->io_size = (file_size *
1109                                               o->size_percent) / 100;
1110                                 if (f->io_size > (file_size - f->file_offset))
1111                                         f->io_size = file_size - f->file_offset;
1112
1113                                 f->io_size -= (f->io_size % td_min_bs(td));
1114                         }
1115                         total_size += f->io_size;
1116                 }
1117
1118                 if (f->filetype == FIO_TYPE_FILE &&
1119                     (f->io_size + f->file_offset) > f->real_file_size) {
1120                         if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1121                             !o->create_on_open) {
1122                                 need_extend++;
1123                                 extend_size += (f->io_size + f->file_offset);
1124                                 fio_file_set_extend(f);
1125                         } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1126                                    (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1127                                     td_ioengine_flagged(td, FIO_FAKEIO)))
1128                                 f->real_file_size = f->io_size + f->file_offset;
1129                 }
1130         }
1131
1132         if (td->o.block_error_hist) {
1133                 int len;
1134
1135                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1136                 f = td->files[0];
1137                 len = f->io_size / td->o.bs[DDIR_TRIM];
1138                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1139                         log_err("fio: cannot calculate block histogram with "
1140                                 "%d trim blocks, maximum %d\n",
1141                                 len, MAX_NR_BLOCK_INFOS);
1142                         td_verror(td, EINVAL, "block_error_hist");
1143                         goto err_out;
1144                 }
1145
1146                 td->ts.nr_block_infos = len;
1147                 for (i = 0; i < len; i++)
1148                         td->ts.block_infos[i] =
1149                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1150         } else
1151                 td->ts.nr_block_infos = 0;
1152
1153         if (!o->size || (total_size && o->size > total_size))
1154                 o->size = total_size;
1155
1156         if (o->size < td_min_bs(td)) {
1157                 log_err("fio: blocksize too large for data set\n");
1158                 goto err_out;
1159         }
1160
1161         /*
1162          * See if we need to extend some files, typically needed when our
1163          * target regular files don't exist yet, but our jobs require them
1164          * initially due to read I/Os.
1165          */
1166         if (need_extend) {
1167                 temp_stall_ts = 1;
1168                 if (output_format & FIO_OUTPUT_NORMAL) {
1169                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1170                                  o->name,
1171                                  need_extend > 1 ? "s" : "",
1172                                  need_extend,
1173                                  need_extend > 1 ? "s" : "",
1174                                  need_extend > 1 ? "total " : "",
1175                                  extend_size >> 20);
1176                 }
1177
1178                 for_each_file(td, f, i) {
1179                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1180
1181                         if (!fio_file_extend(f))
1182                                 continue;
1183
1184                         assert(f->filetype == FIO_TYPE_FILE);
1185                         fio_file_clear_extend(f);
1186                         if (!o->fill_device) {
1187                                 old_len = f->real_file_size;
1188                                 extend_len = f->io_size + f->file_offset -
1189                                                 old_len;
1190                         }
1191                         f->real_file_size = (f->io_size + f->file_offset);
1192                         err = extend_file(td, f);
1193                         if (err)
1194                                 break;
1195
1196                         err = __file_invalidate_cache(td, f, old_len,
1197                                                                 extend_len);
1198
1199                         /*
1200                          * Shut up static checker
1201                          */
1202                         if (f->fd != -1)
1203                                 close(f->fd);
1204
1205                         f->fd = -1;
1206                         if (err)
1207                                 break;
1208                 }
1209                 temp_stall_ts = 0;
1210         }
1211
1212         if (err)
1213                 goto err_out;
1214
1215         /*
1216          * iolog already set the total io size, if we read back
1217          * stored entries.
1218          */
1219         if (!o->read_iolog_file) {
1220                 if (o->io_size)
1221                         td->total_io_size = o->io_size * o->loops;
1222                 else
1223                         td->total_io_size = o->size * o->loops;
1224         }
1225
1226 done:
1227         if (o->create_only)
1228                 td->done = 1;
1229
1230         td_restore_runstate(td, old_state);
1231
1232         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1233                 err = zbd_init(td);
1234                 if (err)
1235                         goto err_out;
1236         }
1237         return 0;
1238
1239 err_offset:
1240         log_err("%s: you need to specify valid offset=\n", o->name);
1241 err_out:
1242         td_restore_runstate(td, old_state);
1243         return 1;
1244 }
1245
1246 bool pre_read_files(struct thread_data *td)
1247 {
1248         struct fio_file *f;
1249         unsigned int i;
1250
1251         dprint(FD_FILE, "pre_read files\n");
1252
1253         for_each_file(td, f, i) {
1254                 if (!pre_read_file(td, f))
1255                         return false;
1256         }
1257
1258         return true;
1259 }
1260
1261 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1262 {
1263         unsigned int range_size, seed;
1264         uint64_t nranges;
1265         uint64_t fsize;
1266
1267         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1268         fsize = min(f->real_file_size, f->io_size);
1269
1270         nranges = (fsize + range_size - 1ULL) / range_size;
1271
1272         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1273         if (!td->o.rand_repeatable)
1274                 seed = td->rand_seeds[4];
1275
1276         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1277                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1278         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1279                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1280         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1281                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1282 }
1283
1284 static bool init_rand_distribution(struct thread_data *td)
1285 {
1286         struct fio_file *f;
1287         unsigned int i;
1288         int state;
1289
1290         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1291                 return false;
1292
1293         state = td_bump_runstate(td, TD_SETTING_UP);
1294
1295         for_each_file(td, f, i)
1296                 __init_rand_distribution(td, f);
1297
1298         td_restore_runstate(td, state);
1299         return true;
1300 }
1301
1302 /*
1303  * Check if the number of blocks exceeds the randomness capability of
1304  * the selected generator. Tausworthe is 32-bit, the others are fullly
1305  * 64-bit capable.
1306  */
1307 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1308                                  uint64_t blocks)
1309 {
1310         if (blocks <= FRAND32_MAX)
1311                 return 0;
1312         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1313                 return 0;
1314
1315         /*
1316          * If the user hasn't specified a random generator, switch
1317          * to tausworthe64 with informational warning. If the user did
1318          * specify one, just warn.
1319          */
1320         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1321                         f->file_name);
1322
1323         if (!fio_option_is_set(&td->o, random_generator)) {
1324                 log_info("fio: Switching to tausworthe64. Use the "
1325                          "random_generator= option to get rid of this "
1326                          "warning.\n");
1327                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1328                 return 0;
1329         }
1330
1331         /*
1332          * Just make this information to avoid breaking scripts.
1333          */
1334         log_info("fio: Use the random_generator= option to switch to lfsr or "
1335                          "tausworthe64.\n");
1336         return 0;
1337 }
1338
1339 bool init_random_map(struct thread_data *td)
1340 {
1341         unsigned long long blocks;
1342         struct fio_file *f;
1343         unsigned int i;
1344
1345         if (init_rand_distribution(td))
1346                 return true;
1347         if (!td_random(td))
1348                 return true;
1349
1350         for_each_file(td, f, i) {
1351                 uint64_t fsize = min(f->real_file_size, f->io_size);
1352
1353                 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1354                         fsize = td->o.zone_range;
1355
1356                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1357
1358                 if (check_rand_gen_limits(td, f, blocks))
1359                         return false;
1360
1361                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1362                         uint64_t seed;
1363
1364                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1365
1366                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1367                                 fio_file_set_lfsr(f);
1368                                 continue;
1369                         }
1370                 } else if (!td->o.norandommap) {
1371                         f->io_axmap = axmap_new(blocks);
1372                         if (f->io_axmap) {
1373                                 fio_file_set_axmap(f);
1374                                 continue;
1375                         }
1376                 } else if (td->o.norandommap)
1377                         continue;
1378
1379                 if (!td->o.softrandommap) {
1380                         log_err("fio: failed allocating random map. If running"
1381                                 " a large number of jobs, try the 'norandommap'"
1382                                 " option or set 'softrandommap'. Or give"
1383                                 " a larger --alloc-size to fio.\n");
1384                         return false;
1385                 }
1386
1387                 log_info("fio: file %s failed allocating random map. Running "
1388                          "job without.\n", f->file_name);
1389         }
1390
1391         return true;
1392 }
1393
1394 void close_files(struct thread_data *td)
1395 {
1396         struct fio_file *f;
1397         unsigned int i;
1398
1399         for_each_file(td, f, i) {
1400                 if (fio_file_open(f))
1401                         td_io_close_file(td, f);
1402         }
1403 }
1404
1405 void close_and_free_files(struct thread_data *td)
1406 {
1407         struct fio_file *f;
1408         unsigned int i;
1409         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1410
1411         dprint(FD_FILE, "close files\n");
1412
1413         for_each_file(td, f, i) {
1414                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1415                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1416                         td_io_unlink_file(td, f);
1417                 }
1418
1419                 if (fio_file_open(f))
1420                         td_io_close_file(td, f);
1421
1422                 remove_file_hash(f);
1423
1424                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1425                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1426                         td_io_unlink_file(td, f);
1427                 }
1428
1429                 zbd_free_zone_info(f);
1430
1431                 if (use_free)
1432                         free(f->file_name);
1433                 else
1434                         sfree(f->file_name);
1435                 f->file_name = NULL;
1436                 if (fio_file_axmap(f)) {
1437                         axmap_free(f->io_axmap);
1438                         f->io_axmap = NULL;
1439                 }
1440                 if (use_free)
1441                         free(f);
1442                 else
1443                         sfree(f);
1444         }
1445
1446         td->o.filename = NULL;
1447         free(td->files);
1448         free(td->file_locks);
1449         td->files_index = 0;
1450         td->files = NULL;
1451         td->file_locks = NULL;
1452         td->o.file_lock_mode = FILE_LOCK_NONE;
1453         td->o.nr_files = 0;
1454 }
1455
1456 static void get_file_type(struct fio_file *f)
1457 {
1458         struct stat sb;
1459
1460         if (!strcmp(f->file_name, "-"))
1461                 f->filetype = FIO_TYPE_PIPE;
1462         else
1463                 f->filetype = FIO_TYPE_FILE;
1464
1465 #ifdef WIN32
1466         /* \\.\ is the device namespace in Windows, where every file is
1467          * a block device */
1468         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1469                 f->filetype = FIO_TYPE_BLOCK;
1470 #endif
1471
1472         if (!stat(f->file_name, &sb)) {
1473                 if (S_ISBLK(sb.st_mode))
1474                         f->filetype = FIO_TYPE_BLOCK;
1475                 else if (S_ISCHR(sb.st_mode))
1476                         f->filetype = FIO_TYPE_CHAR;
1477                 else if (S_ISFIFO(sb.st_mode))
1478                         f->filetype = FIO_TYPE_PIPE;
1479         }
1480 }
1481
1482 static bool __is_already_allocated(const char *fname, bool set)
1483 {
1484         struct flist_head *entry;
1485         bool ret;
1486
1487         ret = file_bloom_exists(fname, set);
1488         if (!ret)
1489                 return ret;
1490
1491         flist_for_each(entry, &filename_list) {
1492                 struct file_name *fn;
1493
1494                 fn = flist_entry(entry, struct file_name, list);
1495
1496                 if (!strcmp(fn->filename, fname))
1497                         return true;
1498         }
1499
1500         return false;
1501 }
1502
1503 static bool is_already_allocated(const char *fname)
1504 {
1505         bool ret;
1506
1507         fio_file_hash_lock();
1508         ret = __is_already_allocated(fname, false);
1509         fio_file_hash_unlock();
1510
1511         return ret;
1512 }
1513
1514 static void set_already_allocated(const char *fname)
1515 {
1516         struct file_name *fn;
1517
1518         fn = malloc(sizeof(struct file_name));
1519         fn->filename = strdup(fname);
1520
1521         fio_file_hash_lock();
1522         if (!__is_already_allocated(fname, true)) {
1523                 flist_add_tail(&fn->list, &filename_list);
1524                 fn = NULL;
1525         }
1526         fio_file_hash_unlock();
1527
1528         if (fn) {
1529                 free(fn->filename);
1530                 free(fn);
1531         }
1532 }
1533
1534 static void free_already_allocated(void)
1535 {
1536         struct flist_head *entry, *tmp;
1537         struct file_name *fn;
1538
1539         if (flist_empty(&filename_list))
1540                 return;
1541
1542         fio_file_hash_lock();
1543         flist_for_each_safe(entry, tmp, &filename_list) {
1544                 fn = flist_entry(entry, struct file_name, list);
1545                 free(fn->filename);
1546                 flist_del(&fn->list);
1547                 free(fn);
1548         }
1549
1550         fio_file_hash_unlock();
1551 }
1552
1553 static struct fio_file *alloc_new_file(struct thread_data *td)
1554 {
1555         struct fio_file *f;
1556
1557         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1558                 f = calloc(1, sizeof(*f));
1559         else
1560                 f = scalloc(1, sizeof(*f));
1561         if (!f) {
1562                 assert(0);
1563                 return NULL;
1564         }
1565
1566         f->fd = -1;
1567         f->shadow_fd = -1;
1568         fio_file_reset(td, f);
1569         return f;
1570 }
1571
1572 bool exists_and_not_regfile(const char *filename)
1573 {
1574         struct stat sb;
1575
1576         if (lstat(filename, &sb) == -1)
1577                 return false;
1578
1579 #ifndef WIN32 /* NOT Windows */
1580         if (S_ISREG(sb.st_mode))
1581                 return false;
1582 #else
1583         /* \\.\ is the device namespace in Windows, where every file
1584          * is a device node */
1585         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1586                 return false;
1587 #endif
1588
1589         return true;
1590 }
1591
1592 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1593 {
1594         int cur_files = td->files_index;
1595         char file_name[PATH_MAX];
1596         struct fio_file *f;
1597         int len = 0;
1598
1599         dprint(FD_FILE, "add file %s\n", fname);
1600
1601         if (td->o.directory)
1602                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1603                                         td->o.unique_filename);
1604
1605         sprintf(file_name + len, "%s", fname);
1606
1607         /* clean cloned siblings using existing files */
1608         if (numjob && is_already_allocated(file_name) &&
1609             !exists_and_not_regfile(fname))
1610                 return 0;
1611
1612         f = alloc_new_file(td);
1613
1614         if (td->files_size <= td->files_index) {
1615                 unsigned int new_size = td->o.nr_files + 1;
1616
1617                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1618
1619                 td->files = realloc(td->files, new_size * sizeof(f));
1620                 if (td->files == NULL) {
1621                         log_err("fio: realloc OOM\n");
1622                         assert(0);
1623                 }
1624                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1625                         td->file_locks = realloc(td->file_locks, new_size);
1626                         if (!td->file_locks) {
1627                                 log_err("fio: realloc OOM\n");
1628                                 assert(0);
1629                         }
1630                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1631                 }
1632                 td->files_size = new_size;
1633         }
1634         td->files[cur_files] = f;
1635         f->fileno = cur_files;
1636
1637         /*
1638          * init function, io engine may not be loaded yet
1639          */
1640         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1641                 f->real_file_size = -1ULL;
1642
1643         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1644                 f->file_name = strdup(file_name);
1645         else
1646                 f->file_name = smalloc_strdup(file_name);
1647
1648         /* can't handle smalloc failure from here */
1649         assert(f->file_name);
1650
1651         get_file_type(f);
1652
1653         switch (td->o.file_lock_mode) {
1654         case FILE_LOCK_NONE:
1655                 break;
1656         case FILE_LOCK_READWRITE:
1657                 f->rwlock = fio_rwlock_init();
1658                 break;
1659         case FILE_LOCK_EXCLUSIVE:
1660                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1661                 break;
1662         default:
1663                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1664                 assert(0);
1665         }
1666
1667         td->files_index++;
1668
1669         if (td->o.numjobs > 1)
1670                 set_already_allocated(file_name);
1671
1672         if (inc)
1673                 td->o.nr_files++;
1674
1675         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1676                                                         cur_files);
1677
1678         return cur_files;
1679 }
1680
1681 int add_file_exclusive(struct thread_data *td, const char *fname)
1682 {
1683         struct fio_file *f;
1684         unsigned int i;
1685
1686         for_each_file(td, f, i) {
1687                 if (!strcmp(f->file_name, fname))
1688                         return i;
1689         }
1690
1691         return add_file(td, fname, 0, 1);
1692 }
1693
1694 void get_file(struct fio_file *f)
1695 {
1696         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1697         assert(fio_file_open(f));
1698         f->references++;
1699 }
1700
1701 int put_file(struct thread_data *td, struct fio_file *f)
1702 {
1703         int f_ret = 0, ret = 0;
1704
1705         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1706
1707         if (!fio_file_open(f)) {
1708                 assert(f->fd == -1);
1709                 return 0;
1710         }
1711
1712         assert(f->references);
1713         if (--f->references)
1714                 return 0;
1715
1716         disk_util_dec(f->du);
1717
1718         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1719                 unlock_file_all(td, f);
1720
1721         if (should_fsync(td) && td->o.fsync_on_close) {
1722                 f_ret = fsync(f->fd);
1723                 if (f_ret < 0)
1724                         f_ret = errno;
1725         }
1726
1727         if (td->io_ops->close_file)
1728                 ret = td->io_ops->close_file(td, f);
1729
1730         if (!ret)
1731                 ret = f_ret;
1732
1733         td->nr_open_files--;
1734         fio_file_clear_closing(f);
1735         fio_file_clear_open(f);
1736         assert(f->fd == -1);
1737         return ret;
1738 }
1739
1740 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1741 {
1742         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1743                 return;
1744
1745         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1746                 if (ddir == DDIR_READ)
1747                         fio_rwlock_read(f->rwlock);
1748                 else
1749                         fio_rwlock_write(f->rwlock);
1750         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1751                 fio_sem_down(f->lock);
1752
1753         td->file_locks[f->fileno] = td->o.file_lock_mode;
1754 }
1755
1756 void unlock_file(struct thread_data *td, struct fio_file *f)
1757 {
1758         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1759                 return;
1760
1761         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1762                 fio_rwlock_unlock(f->rwlock);
1763         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1764                 fio_sem_up(f->lock);
1765
1766         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1767 }
1768
1769 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1770 {
1771         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1772                 return;
1773         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1774                 unlock_file(td, f);
1775 }
1776
1777 static bool recurse_dir(struct thread_data *td, const char *dirname)
1778 {
1779         struct dirent *dir;
1780         bool ret = false;
1781         DIR *D;
1782
1783         D = opendir(dirname);
1784         if (!D) {
1785                 char buf[FIO_VERROR_SIZE];
1786
1787                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1788                 td_verror(td, errno, buf);
1789                 return true;
1790         }
1791
1792         while ((dir = readdir(D)) != NULL) {
1793                 char full_path[PATH_MAX];
1794                 struct stat sb;
1795
1796                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1797                         continue;
1798
1799                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1800
1801                 if (lstat(full_path, &sb) == -1) {
1802                         if (errno != ENOENT) {
1803                                 td_verror(td, errno, "stat");
1804                                 ret = true;
1805                                 break;
1806                         }
1807                 }
1808
1809                 if (S_ISREG(sb.st_mode)) {
1810                         add_file(td, full_path, 0, 1);
1811                         continue;
1812                 }
1813                 if (!S_ISDIR(sb.st_mode))
1814                         continue;
1815
1816                 ret = recurse_dir(td, full_path);
1817                 if (ret)
1818                         break;
1819         }
1820
1821         closedir(D);
1822         return ret;
1823 }
1824
1825 int add_dir_files(struct thread_data *td, const char *path)
1826 {
1827         int ret = recurse_dir(td, path);
1828
1829         if (!ret)
1830                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1831
1832         return ret;
1833 }
1834
1835 void dup_files(struct thread_data *td, struct thread_data *org)
1836 {
1837         struct fio_file *f;
1838         unsigned int i;
1839
1840         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1841
1842         if (!org->files)
1843                 return;
1844
1845         td->files = malloc(org->files_index * sizeof(f));
1846
1847         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1848                 td->file_locks = malloc(org->files_index);
1849
1850         for_each_file(org, f, i) {
1851                 struct fio_file *__f;
1852
1853                 __f = alloc_new_file(td);
1854
1855                 if (f->file_name) {
1856                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1857                                 __f->file_name = strdup(f->file_name);
1858                         else
1859                                 __f->file_name = smalloc_strdup(f->file_name);
1860
1861                         /* can't handle smalloc failure from here */
1862                         assert(__f->file_name);
1863                         __f->filetype = f->filetype;
1864                 }
1865
1866                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1867                         __f->lock = f->lock;
1868                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1869                         __f->rwlock = f->rwlock;
1870
1871                 td->files[i] = __f;
1872         }
1873 }
1874
1875 /*
1876  * Returns the index that matches the filename, or -1 if not there
1877  */
1878 int get_fileno(struct thread_data *td, const char *fname)
1879 {
1880         struct fio_file *f;
1881         unsigned int i;
1882
1883         for_each_file(td, f, i)
1884                 if (!strcmp(f->file_name, fname))
1885                         return i;
1886
1887         return -1;
1888 }
1889
1890 /*
1891  * For log usage, where we add/open/close files automatically
1892  */
1893 void free_release_files(struct thread_data *td)
1894 {
1895         close_files(td);
1896         td->o.nr_files = 0;
1897         td->o.open_files = 0;
1898         td->files_index = 0;
1899 }
1900
1901 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1902 {
1903         int i;
1904
1905         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1906                 f->last_pos[i] = f->file_offset;
1907                 f->last_start[i] = -1ULL;
1908         }
1909
1910         if (fio_file_axmap(f))
1911                 axmap_reset(f->io_axmap);
1912         else if (fio_file_lfsr(f))
1913                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1914
1915         zbd_file_reset(td, f);
1916 }
1917
1918 bool fio_files_done(struct thread_data *td)
1919 {
1920         struct fio_file *f;
1921         unsigned int i;
1922
1923         for_each_file(td, f, i)
1924                 if (!fio_file_done(f))
1925                         return false;
1926
1927         return true;
1928 }
1929
1930 /* free memory used in initialization phase only */
1931 void filesetup_mem_free(void)
1932 {
1933         free_already_allocated();
1934 }
1935
1936 /*
1937  * This function is for platforms which support direct I/O but not O_DIRECT.
1938  */
1939 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1940 {
1941 #ifdef FIO_OS_DIRECTIO
1942         int ret = fio_set_odirect(f);
1943
1944         if (ret) {
1945                 td_verror(td, ret, "fio_set_directio");
1946 #if defined(__sun__)
1947                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1948                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1949                 } else {
1950                         log_err("fio: the file system does not seem to support direct IO\n");
1951                 }
1952 #else
1953                 log_err("fio: the file system does not seem to support direct IO\n");
1954 #endif
1955                 return -1;
1956         }
1957
1958         return 0;
1959 #else
1960         log_err("fio: direct IO is not supported on this host operating system\n");
1961         return -1;
1962 #endif
1963 }