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 /*
 917  * Find longest path component that exists and return its length
 918  */
 919 int longest_existing_path(char *path) {
 920         char buf[PATH_MAX];
 921         bool done;
 922         char *buf_pos;
 923         int offset;
 924 #ifdef WIN32
 925         DWORD dwAttr;
 926 #else
 927         struct stat sb;
 928 #endif
 929
 930         sprintf(buf, "%s", path);
 931         done = false;
 932         while (!done) {
 933                 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
 934                 if (!buf_pos) {
 935                         done = true;
 936                         offset = 0;
 937                         break;
 938                 }
 939
 940                 *(buf_pos + 1) = '\0';
 941
 942 #ifdef WIN32
 943                 dwAttr = GetFileAttributesA(buf);
 944                 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
 945                         done = true;
 946                 }
 947 #else
 948                 if (stat(buf, &sb) == 0)
 949                         done = true;
 950 #endif
 951                 if (done)
 952                         offset = buf_pos - buf;
 953                 else
 954                         *buf_pos = '\0';
 955         }
 956
 957         return offset;
 958 }
 959
 960 static bool create_work_dirs(struct thread_data *td, const char *fname)
 961 {
 962         char path[PATH_MAX];
 963         char *start, *end;
 964         int offset;
 965
 966         snprintf(path, PATH_MAX, "%s", fname);
 967         start = path;
 968
 969         offset = longest_existing_path(path);
 970         end = start + offset;
 971         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
 972                 if (end == start) {
 973                         end++;
 974                         continue;
 975                 }
 976                 *end = '\0';
 977                 errno = 0;
 978                 if (fio_mkdir(path, 0700) && errno != EEXIST) {
 979                         log_err("fio: failed to create dir (%s): %s\n",
 980                                 start, strerror(errno));
 981                         return false;
 982                 }
 983                 *end = FIO_OS_PATH_SEPARATOR;
 984                 end++;
 985         }
 986         td->flags |= TD_F_DIRS_CREATED;
 987         return true;
 988 }
 989
 990 /*
 991  * Open the files and setup files sizes, creating files if necessary.
 992  */
 993 int setup_files(struct thread_data *td)
 994 {
 995         unsigned long long total_size, extend_size;
 996         struct thread_options *o = &td->o;
 997         struct fio_file *f;
 998         unsigned int i, nr_fs_extra = 0;
 999         int err = 0, need_extend;
1000         int old_state;
1001         const unsigned long long bs = td_min_bs(td);
1002         uint64_t fs = 0;
1003
1004         dprint(FD_FILE, "setup files\n");
1005
1006         old_state = td_bump_runstate(td, TD_SETTING_UP);
1007
1008         for_each_file(td, f, i) {
1009                 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1010                     strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1011                     !(td->flags & TD_F_DIRS_CREATED) &&
1012                     !create_work_dirs(td, f->file_name))
1013                         goto err_out;
1014         }
1015
1016         /*
1017          * Find out physical size of files or devices for this thread,
1018          * before we determine I/O size and range of our targets.
1019          * If ioengine defines a setup() method, it's responsible for
1020          * opening the files and setting f->real_file_size to indicate
1021          * the valid range for that file.
1022          */
1023         if (td->io_ops->setup)
1024                 err = td->io_ops->setup(td);
1025         else
1026                 err = get_file_sizes(td);
1027
1028         if (err)
1029                 goto err_out;
1030
1031         if (o->read_iolog_file)
1032                 goto done;
1033
1034         /*
1035          * check sizes. if the files/devices do not exist and the size
1036          * isn't passed to fio, abort.
1037          */
1038         total_size = 0;
1039         for_each_file(td, f, i) {
1040                 f->fileno = i;
1041                 if (f->real_file_size == -1ULL)
1042                         total_size = -1ULL;
1043                 else
1044                         total_size += f->real_file_size;
1045         }
1046
1047         if (o->fill_device)
1048                 td->fill_device_size = get_fs_free_counts(td);
1049
1050         /*
1051          * device/file sizes are zero and no size given, punt
1052          */
1053         if ((!total_size || total_size == -1ULL) && !o->size &&
1054             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1055             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1056                 log_err("%s: you need to specify size=\n", o->name);
1057                 td_verror(td, EINVAL, "total_file_size");
1058                 goto err_out;
1059         }
1060
1061         /*
1062          * Calculate per-file size and potential extra size for the
1063          * first files, if needed (i.e. if we don't have a fixed size).
1064          */
1065         if (!o->file_size_low && o->nr_files) {
1066                 uint64_t all_fs;
1067
1068                 fs = o->size / o->nr_files;
1069                 all_fs = fs * o->nr_files;
1070
1071                 if (all_fs < o->size)
1072                         nr_fs_extra = (o->size - all_fs) / bs;
1073         }
1074
1075         /*
1076          * now file sizes are known, so we can set ->io_size. if size= is
1077          * not given, ->io_size is just equal to ->real_file_size. if size
1078          * is given, ->io_size is size / nr_files.
1079          */
1080         extend_size = total_size = 0;
1081         need_extend = 0;
1082         for_each_file(td, f, i) {
1083                 f->file_offset = get_start_offset(td, f);
1084
1085                 /*
1086                  * Update ->io_size depending on options specified.
1087                  * ->file_size_low being 0 means filesize option isn't set.
1088                  * Non zero ->file_size_low equals ->file_size_high means
1089                  * filesize option is set in a fixed size format.
1090                  * Non zero ->file_size_low not equals ->file_size_high means
1091                  * filesize option is set in a range format.
1092                  */
1093                 if (!o->file_size_low) {
1094                         /*
1095                          * no file size or range given, file size is equal to
1096                          * total size divided by number of files. If the size
1097                          * doesn't divide nicely with the min blocksize,
1098                          * make the first files bigger.
1099                          */
1100                         f->io_size = fs;
1101                         if (nr_fs_extra) {
1102                                 nr_fs_extra--;
1103                                 f->io_size += bs;
1104                         }
1105
1106                         /*
1107                          * We normally don't come here for regular files, but
1108                          * if the result is 0 for a regular file, set it to the
1109                          * real file size. This could be size of the existing
1110                          * one if it already exists, but otherwise will be set
1111                          * to 0. A new file won't be created because
1112                          * ->io_size + ->file_offset equals ->real_file_size.
1113                          */
1114                         if (!f->io_size) {
1115                                 if (f->file_offset > f->real_file_size)
1116                                         goto err_offset;
1117                                 f->io_size = f->real_file_size - f->file_offset;
1118                                 if (!f->io_size)
1119                                         log_info("fio: file %s may be ignored\n",
1120                                                 f->file_name);
1121                         }
1122                 } else if (f->real_file_size < o->file_size_low ||
1123                            f->real_file_size > o->file_size_high) {
1124                         if (f->file_offset > o->file_size_low)
1125                                 goto err_offset;
1126                         /*
1127                          * file size given. if it's fixed, use that. if it's a
1128                          * range, generate a random size in-between.
1129                          */
1130                         if (o->file_size_low == o->file_size_high)
1131                                 f->io_size = o->file_size_low - f->file_offset;
1132                         else {
1133                                 f->io_size = get_rand_file_size(td)
1134                                                 - f->file_offset;
1135                         }
1136                 } else
1137                         f->io_size = f->real_file_size - f->file_offset;
1138
1139                 if (f->io_size == -1ULL)
1140                         total_size = -1ULL;
1141                 else {
1142                         uint64_t io_size;
1143
1144                         if (o->size_percent && o->size_percent != 100) {
1145                                 uint64_t file_size;
1146
1147                                 file_size = f->io_size + f->file_offset;
1148                                 f->io_size = (file_size *
1149                                               o->size_percent) / 100;
1150                                 if (f->io_size > (file_size - f->file_offset))
1151                                         f->io_size = file_size - f->file_offset;
1152
1153                                 f->io_size -= (f->io_size % td_min_bs(td));
1154                         }
1155
1156                         io_size = f->io_size;
1157                         if (o->io_size_percent && o->io_size_percent != 100) {
1158                                 io_size *= o->io_size_percent;
1159                                 io_size /= 100;
1160                         }
1161
1162                         total_size += io_size;
1163                 }
1164
1165                 if (f->filetype == FIO_TYPE_FILE &&
1166                     (f->io_size + f->file_offset) > f->real_file_size) {
1167                         if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1168                             !o->create_on_open) {
1169                                 need_extend++;
1170                                 extend_size += (f->io_size + f->file_offset);
1171                                 fio_file_set_extend(f);
1172                         } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1173                                    (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1174                                     td_ioengine_flagged(td, FIO_FAKEIO)))
1175                                 f->real_file_size = f->io_size + f->file_offset;
1176                 }
1177         }
1178
1179         if (td->o.block_error_hist) {
1180                 int len;
1181
1182                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1183                 f = td->files[0];
1184                 len = f->io_size / td->o.bs[DDIR_TRIM];
1185                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1186                         log_err("fio: cannot calculate block histogram with "
1187                                 "%d trim blocks, maximum %d\n",
1188                                 len, MAX_NR_BLOCK_INFOS);
1189                         td_verror(td, EINVAL, "block_error_hist");
1190                         goto err_out;
1191                 }
1192
1193                 td->ts.nr_block_infos = len;
1194                 for (i = 0; i < len; i++)
1195                         td->ts.block_infos[i] =
1196                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1197         } else
1198                 td->ts.nr_block_infos = 0;
1199
1200         if (!o->size || (total_size && o->size > total_size))
1201                 o->size = total_size;
1202
1203         if (o->size < td_min_bs(td)) {
1204                 log_err("fio: blocksize too large for data set\n");
1205                 goto err_out;
1206         }
1207
1208         /*
1209          * See if we need to extend some files, typically needed when our
1210          * target regular files don't exist yet, but our jobs require them
1211          * initially due to read I/Os.
1212          */
1213         if (need_extend) {
1214                 temp_stall_ts = 1;
1215                 if (output_format & FIO_OUTPUT_NORMAL) {
1216                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1217                                  o->name,
1218                                  need_extend > 1 ? "s" : "",
1219                                  need_extend,
1220                                  need_extend > 1 ? "s" : "",
1221                                  need_extend > 1 ? "total " : "",
1222                                  extend_size >> 20);
1223                 }
1224
1225                 for_each_file(td, f, i) {
1226                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1227
1228                         if (!fio_file_extend(f))
1229                                 continue;
1230
1231                         assert(f->filetype == FIO_TYPE_FILE);
1232                         fio_file_clear_extend(f);
1233                         if (!o->fill_device) {
1234                                 old_len = f->real_file_size;
1235                                 extend_len = f->io_size + f->file_offset -
1236                                                 old_len;
1237                         }
1238                         f->real_file_size = (f->io_size + f->file_offset);
1239                         err = extend_file(td, f);
1240                         if (err)
1241                                 break;
1242
1243                         err = __file_invalidate_cache(td, f, old_len,
1244                                                                 extend_len);
1245
1246                         /*
1247                          * Shut up static checker
1248                          */
1249                         if (f->fd != -1)
1250                                 close(f->fd);
1251
1252                         f->fd = -1;
1253                         if (err)
1254                                 break;
1255                 }
1256                 temp_stall_ts = 0;
1257         }
1258
1259         if (err)
1260                 goto err_out;
1261
1262         /*
1263          * iolog already set the total io size, if we read back
1264          * stored entries.
1265          */
1266         if (!o->read_iolog_file) {
1267                 if (o->io_size)
1268                         td->total_io_size = o->io_size * o->loops;
1269                 else
1270                         td->total_io_size = o->size * o->loops;
1271         }
1272
1273 done:
1274         if (o->create_only)
1275                 td->done = 1;
1276
1277         td_restore_runstate(td, old_state);
1278
1279         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1280                 err = zbd_setup_files(td);
1281                 if (err)
1282                         goto err_out;
1283         }
1284         return 0;
1285
1286 err_offset:
1287         log_err("%s: you need to specify valid offset=\n", o->name);
1288 err_out:
1289         td_restore_runstate(td, old_state);
1290         return 1;
1291 }
1292
1293 bool pre_read_files(struct thread_data *td)
1294 {
1295         struct fio_file *f;
1296         unsigned int i;
1297
1298         dprint(FD_FILE, "pre_read files\n");
1299
1300         for_each_file(td, f, i) {
1301                 if (!pre_read_file(td, f))
1302                         return false;
1303         }
1304
1305         return true;
1306 }
1307
1308 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1309 {
1310         unsigned int range_size, seed;
1311         uint64_t nranges;
1312         uint64_t fsize;
1313
1314         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1315         fsize = min(f->real_file_size, f->io_size);
1316
1317         nranges = (fsize + range_size - 1ULL) / range_size;
1318
1319         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1320         if (!td->o.rand_repeatable)
1321                 seed = td->rand_seeds[4];
1322
1323         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1324                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1325         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1326                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1327         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1328                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1329 }
1330
1331 static bool init_rand_distribution(struct thread_data *td)
1332 {
1333         struct fio_file *f;
1334         unsigned int i;
1335         int state;
1336
1337         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1338             td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1339             td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1340                 return false;
1341
1342         state = td_bump_runstate(td, TD_SETTING_UP);
1343
1344         for_each_file(td, f, i)
1345                 __init_rand_distribution(td, f);
1346
1347         td_restore_runstate(td, state);
1348         return true;
1349 }
1350
1351 /*
1352  * Check if the number of blocks exceeds the randomness capability of
1353  * the selected generator. Tausworthe is 32-bit, the others are fullly
1354  * 64-bit capable.
1355  */
1356 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1357                                  uint64_t blocks)
1358 {
1359         if (blocks <= FRAND32_MAX)
1360                 return 0;
1361         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1362                 return 0;
1363
1364         /*
1365          * If the user hasn't specified a random generator, switch
1366          * to tausworthe64 with informational warning. If the user did
1367          * specify one, just warn.
1368          */
1369         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1370                         f->file_name);
1371
1372         if (!fio_option_is_set(&td->o, random_generator)) {
1373                 log_info("fio: Switching to tausworthe64. Use the "
1374                          "random_generator= option to get rid of this "
1375                          "warning.\n");
1376                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1377                 return 0;
1378         }
1379
1380         /*
1381          * Just make this information to avoid breaking scripts.
1382          */
1383         log_info("fio: Use the random_generator= option to switch to lfsr or "
1384                          "tausworthe64.\n");
1385         return 0;
1386 }
1387
1388 bool init_random_map(struct thread_data *td)
1389 {
1390         unsigned long long blocks;
1391         struct fio_file *f;
1392         unsigned int i;
1393
1394         if (init_rand_distribution(td))
1395                 return true;
1396         if (!td_random(td))
1397                 return true;
1398
1399         for_each_file(td, f, i) {
1400                 uint64_t fsize = min(f->real_file_size, f->io_size);
1401
1402                 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1403                         fsize = td->o.zone_range;
1404
1405                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1406
1407                 if (check_rand_gen_limits(td, f, blocks))
1408                         return false;
1409
1410                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1411                         uint64_t seed;
1412
1413                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1414
1415                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1416                                 fio_file_set_lfsr(f);
1417                                 continue;
1418                         } else {
1419                                 log_err("fio: failed initializing LFSR\n");
1420                                 return false;
1421                         }
1422                 } else if (!td->o.norandommap) {
1423                         f->io_axmap = axmap_new(blocks);
1424                         if (f->io_axmap) {
1425                                 fio_file_set_axmap(f);
1426                                 continue;
1427                         }
1428                 } else if (td->o.norandommap)
1429                         continue;
1430
1431                 if (!td->o.softrandommap) {
1432                         log_err("fio: failed allocating random map. If running"
1433                                 " a large number of jobs, try the 'norandommap'"
1434                                 " option or set 'softrandommap'. Or give"
1435                                 " a larger --alloc-size to fio.\n");
1436                         return false;
1437                 }
1438
1439                 log_info("fio: file %s failed allocating random map. Running "
1440                          "job without.\n", f->file_name);
1441         }
1442
1443         return true;
1444 }
1445
1446 void close_files(struct thread_data *td)
1447 {
1448         struct fio_file *f;
1449         unsigned int i;
1450
1451         for_each_file(td, f, i) {
1452                 if (fio_file_open(f))
1453                         td_io_close_file(td, f);
1454         }
1455 }
1456
1457 void fio_file_free(struct fio_file *f)
1458 {
1459         if (fio_file_axmap(f))
1460                 axmap_free(f->io_axmap);
1461         if (!fio_file_smalloc(f)) {
1462                 free(f->file_name);
1463                 free(f);
1464         } else {
1465                 sfree(f->file_name);
1466                 sfree(f);
1467         }
1468 }
1469
1470 void close_and_free_files(struct thread_data *td)
1471 {
1472         struct fio_file *f;
1473         unsigned int i;
1474
1475         dprint(FD_FILE, "close files\n");
1476
1477         for_each_file(td, f, i) {
1478                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1479                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1480                         td_io_unlink_file(td, f);
1481                 }
1482
1483                 if (fio_file_open(f))
1484                         td_io_close_file(td, f);
1485
1486                 remove_file_hash(f);
1487
1488                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1489                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1490                         td_io_unlink_file(td, f);
1491                 }
1492
1493                 zbd_close_file(f);
1494                 fio_file_free(f);
1495         }
1496
1497         td->o.filename = NULL;
1498         free(td->files);
1499         free(td->file_locks);
1500         td->files_index = 0;
1501         td->files = NULL;
1502         td->file_locks = NULL;
1503         td->o.file_lock_mode = FILE_LOCK_NONE;
1504         td->o.nr_files = 0;
1505 }
1506
1507 static void get_file_type(struct fio_file *f)
1508 {
1509         struct stat sb;
1510
1511         if (!strcmp(f->file_name, "-"))
1512                 f->filetype = FIO_TYPE_PIPE;
1513         else
1514                 f->filetype = FIO_TYPE_FILE;
1515
1516 #ifdef WIN32
1517         /* \\.\ is the device namespace in Windows, where every file is
1518          * a block device */
1519         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1520                 f->filetype = FIO_TYPE_BLOCK;
1521 #endif
1522
1523         if (!stat(f->file_name, &sb)) {
1524                 if (S_ISBLK(sb.st_mode))
1525                         f->filetype = FIO_TYPE_BLOCK;
1526                 else if (S_ISCHR(sb.st_mode))
1527                         f->filetype = FIO_TYPE_CHAR;
1528                 else if (S_ISFIFO(sb.st_mode))
1529                         f->filetype = FIO_TYPE_PIPE;
1530         }
1531 }
1532
1533 static bool __is_already_allocated(const char *fname, bool set)
1534 {
1535         struct flist_head *entry;
1536         bool ret;
1537
1538         ret = file_bloom_exists(fname, set);
1539         if (!ret)
1540                 return ret;
1541
1542         flist_for_each(entry, &filename_list) {
1543                 struct file_name *fn;
1544
1545                 fn = flist_entry(entry, struct file_name, list);
1546
1547                 if (!strcmp(fn->filename, fname))
1548                         return true;
1549         }
1550
1551         return false;
1552 }
1553
1554 static bool is_already_allocated(const char *fname)
1555 {
1556         bool ret;
1557
1558         fio_file_hash_lock();
1559         ret = __is_already_allocated(fname, false);
1560         fio_file_hash_unlock();
1561
1562         return ret;
1563 }
1564
1565 static void set_already_allocated(const char *fname)
1566 {
1567         struct file_name *fn;
1568
1569         fn = malloc(sizeof(struct file_name));
1570         fn->filename = strdup(fname);
1571
1572         fio_file_hash_lock();
1573         if (!__is_already_allocated(fname, true)) {
1574                 flist_add_tail(&fn->list, &filename_list);
1575                 fn = NULL;
1576         }
1577         fio_file_hash_unlock();
1578
1579         if (fn) {
1580                 free(fn->filename);
1581                 free(fn);
1582         }
1583 }
1584
1585 static void free_already_allocated(void)
1586 {
1587         struct flist_head *entry, *tmp;
1588         struct file_name *fn;
1589
1590         if (flist_empty(&filename_list))
1591                 return;
1592
1593         fio_file_hash_lock();
1594         flist_for_each_safe(entry, tmp, &filename_list) {
1595                 fn = flist_entry(entry, struct file_name, list);
1596                 free(fn->filename);
1597                 flist_del(&fn->list);
1598                 free(fn);
1599         }
1600
1601         fio_file_hash_unlock();
1602 }
1603
1604 static struct fio_file *alloc_new_file(struct thread_data *td)
1605 {
1606         struct fio_file *f;
1607
1608         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1609                 f = calloc(1, sizeof(*f));
1610         else
1611                 f = scalloc(1, sizeof(*f));
1612         if (!f) {
1613                 assert(0);
1614                 return NULL;
1615         }
1616
1617         f->fd = -1;
1618         f->shadow_fd = -1;
1619         fio_file_reset(td, f);
1620         if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1621                 fio_file_set_smalloc(f);
1622         return f;
1623 }
1624
1625 bool exists_and_not_regfile(const char *filename)
1626 {
1627         struct stat sb;
1628
1629         if (lstat(filename, &sb) == -1)
1630                 return false;
1631
1632 #ifndef WIN32 /* NOT Windows */
1633         if (S_ISREG(sb.st_mode))
1634                 return false;
1635 #else
1636         /* \\.\ is the device namespace in Windows, where every file
1637          * is a device node */
1638         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1639                 return false;
1640 #endif
1641
1642         return true;
1643 }
1644
1645 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1646 {
1647         int cur_files = td->files_index;
1648         char file_name[PATH_MAX];
1649         struct fio_file *f;
1650         int len = 0;
1651
1652         dprint(FD_FILE, "add file %s\n", fname);
1653
1654         if (td->o.directory)
1655                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1656                                         td->o.unique_filename);
1657
1658         sprintf(file_name + len, "%s", fname);
1659
1660         /* clean cloned siblings using existing files */
1661         if (numjob && is_already_allocated(file_name) &&
1662             !exists_and_not_regfile(fname))
1663                 return 0;
1664
1665         f = alloc_new_file(td);
1666
1667         if (td->files_size <= td->files_index) {
1668                 unsigned int new_size = td->o.nr_files + 1;
1669
1670                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1671
1672                 td->files = realloc(td->files, new_size * sizeof(f));
1673                 if (td->files == NULL) {
1674                         log_err("fio: realloc OOM\n");
1675                         assert(0);
1676                 }
1677                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1678                         td->file_locks = realloc(td->file_locks, new_size);
1679                         if (!td->file_locks) {
1680                                 log_err("fio: realloc OOM\n");
1681                                 assert(0);
1682                         }
1683                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1684                 }
1685                 td->files_size = new_size;
1686         }
1687         td->files[cur_files] = f;
1688         f->fileno = cur_files;
1689
1690         /*
1691          * init function, io engine may not be loaded yet
1692          */
1693         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1694                 f->real_file_size = -1ULL;
1695
1696         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1697                 f->file_name = strdup(file_name);
1698         else
1699                 f->file_name = smalloc_strdup(file_name);
1700
1701         /* can't handle smalloc failure from here */
1702         assert(f->file_name);
1703
1704         get_file_type(f);
1705
1706         switch (td->o.file_lock_mode) {
1707         case FILE_LOCK_NONE:
1708                 break;
1709         case FILE_LOCK_READWRITE:
1710                 f->rwlock = fio_rwlock_init();
1711                 break;
1712         case FILE_LOCK_EXCLUSIVE:
1713                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1714                 break;
1715         default:
1716                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1717                 assert(0);
1718         }
1719
1720         td->files_index++;
1721
1722         if (td->o.numjobs > 1)
1723                 set_already_allocated(file_name);
1724
1725         if (inc)
1726                 td->o.nr_files++;
1727
1728         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1729                                                         cur_files);
1730
1731         return cur_files;
1732 }
1733
1734 int add_file_exclusive(struct thread_data *td, const char *fname)
1735 {
1736         struct fio_file *f;
1737         unsigned int i;
1738
1739         for_each_file(td, f, i) {
1740                 if (!strcmp(f->file_name, fname))
1741                         return i;
1742         }
1743
1744         return add_file(td, fname, 0, 1);
1745 }
1746
1747 void get_file(struct fio_file *f)
1748 {
1749         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1750         assert(fio_file_open(f));
1751         f->references++;
1752 }
1753
1754 int put_file(struct thread_data *td, struct fio_file *f)
1755 {
1756         int f_ret = 0, ret = 0;
1757
1758         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1759
1760         if (!fio_file_open(f)) {
1761                 assert(f->fd == -1);
1762                 return 0;
1763         }
1764
1765         assert(f->references);
1766         if (--f->references)
1767                 return 0;
1768
1769         disk_util_dec(f->du);
1770
1771         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1772                 unlock_file_all(td, f);
1773
1774         if (should_fsync(td) && td->o.fsync_on_close) {
1775                 f_ret = fsync(f->fd);
1776                 if (f_ret < 0)
1777                         f_ret = errno;
1778         }
1779
1780         if (td->io_ops->close_file)
1781                 ret = td->io_ops->close_file(td, f);
1782
1783         if (!ret)
1784                 ret = f_ret;
1785
1786         td->nr_open_files--;
1787         fio_file_clear_closing(f);
1788         fio_file_clear_open(f);
1789         assert(f->fd == -1);
1790         return ret;
1791 }
1792
1793 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1794 {
1795         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1796                 return;
1797
1798         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1799                 if (ddir == DDIR_READ)
1800                         fio_rwlock_read(f->rwlock);
1801                 else
1802                         fio_rwlock_write(f->rwlock);
1803         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1804                 fio_sem_down(f->lock);
1805
1806         td->file_locks[f->fileno] = td->o.file_lock_mode;
1807 }
1808
1809 void unlock_file(struct thread_data *td, struct fio_file *f)
1810 {
1811         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1812                 return;
1813
1814         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1815                 fio_rwlock_unlock(f->rwlock);
1816         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1817                 fio_sem_up(f->lock);
1818
1819         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1820 }
1821
1822 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1823 {
1824         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1825                 return;
1826         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1827                 unlock_file(td, f);
1828 }
1829
1830 static bool recurse_dir(struct thread_data *td, const char *dirname)
1831 {
1832         struct dirent *dir;
1833         bool ret = false;
1834         DIR *D;
1835
1836         D = opendir(dirname);
1837         if (!D) {
1838                 char buf[FIO_VERROR_SIZE];
1839
1840                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1841                 td_verror(td, errno, buf);
1842                 return true;
1843         }
1844
1845         while ((dir = readdir(D)) != NULL) {
1846                 char full_path[PATH_MAX];
1847                 struct stat sb;
1848
1849                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1850                         continue;
1851
1852                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1853
1854                 if (lstat(full_path, &sb) == -1) {
1855                         if (errno != ENOENT) {
1856                                 td_verror(td, errno, "stat");
1857                                 ret = true;
1858                                 break;
1859                         }
1860                 }
1861
1862                 if (S_ISREG(sb.st_mode)) {
1863                         add_file(td, full_path, 0, 1);
1864                         continue;
1865                 }
1866                 if (!S_ISDIR(sb.st_mode))
1867                         continue;
1868
1869                 ret = recurse_dir(td, full_path);
1870                 if (ret)
1871                         break;
1872         }
1873
1874         closedir(D);
1875         return ret;
1876 }
1877
1878 int add_dir_files(struct thread_data *td, const char *path)
1879 {
1880         int ret = recurse_dir(td, path);
1881
1882         if (!ret)
1883                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1884
1885         return ret;
1886 }
1887
1888 void dup_files(struct thread_data *td, struct thread_data *org)
1889 {
1890         struct fio_file *f;
1891         unsigned int i;
1892
1893         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1894
1895         if (!org->files)
1896                 return;
1897
1898         td->files = malloc(org->files_index * sizeof(f));
1899
1900         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1901                 td->file_locks = malloc(org->files_index);
1902
1903         for_each_file(org, f, i) {
1904                 struct fio_file *__f;
1905
1906                 __f = alloc_new_file(td);
1907
1908                 if (f->file_name) {
1909                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1910                                 __f->file_name = strdup(f->file_name);
1911                         else
1912                                 __f->file_name = smalloc_strdup(f->file_name);
1913
1914                         /* can't handle smalloc failure from here */
1915                         assert(__f->file_name);
1916                         __f->filetype = f->filetype;
1917                 }
1918
1919                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1920                         __f->lock = f->lock;
1921                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1922                         __f->rwlock = f->rwlock;
1923
1924                 td->files[i] = __f;
1925         }
1926 }
1927
1928 /*
1929  * Returns the index that matches the filename, or -1 if not there
1930  */
1931 int get_fileno(struct thread_data *td, const char *fname)
1932 {
1933         struct fio_file *f;
1934         unsigned int i;
1935
1936         for_each_file(td, f, i)
1937                 if (!strcmp(f->file_name, fname))
1938                         return i;
1939
1940         return -1;
1941 }
1942
1943 /*
1944  * For log usage, where we add/open/close files automatically
1945  */
1946 void free_release_files(struct thread_data *td)
1947 {
1948         close_files(td);
1949         td->o.nr_files = 0;
1950         td->o.open_files = 0;
1951         td->files_index = 0;
1952 }
1953
1954 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1955 {
1956         int i;
1957
1958         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1959                 f->last_pos[i] = f->file_offset;
1960                 f->last_start[i] = -1ULL;
1961         }
1962
1963         if (fio_file_axmap(f))
1964                 axmap_reset(f->io_axmap);
1965         else if (fio_file_lfsr(f))
1966                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1967
1968         zbd_file_reset(td, f);
1969 }
1970
1971 bool fio_files_done(struct thread_data *td)
1972 {
1973         struct fio_file *f;
1974         unsigned int i;
1975
1976         for_each_file(td, f, i)
1977                 if (!fio_file_done(f))
1978                         return false;
1979
1980         return true;
1981 }
1982
1983 /* free memory used in initialization phase only */
1984 void filesetup_mem_free(void)
1985 {
1986         free_already_allocated();
1987 }
1988
1989 /*
1990  * This function is for platforms which support direct I/O but not O_DIRECT.
1991  */
1992 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1993 {
1994 #ifdef FIO_OS_DIRECTIO
1995         int ret = fio_set_odirect(f);
1996
1997         if (ret) {
1998                 td_verror(td, ret, "fio_set_directio");
1999 #if defined(__sun__)
2000                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2001                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2002                 } else {
2003                         log_err("fio: the file system does not seem to support direct IO\n");
2004                 }
2005 #else
2006                 log_err("fio: the file system does not seem to support direct IO\n");
2007 #endif
2008                 return -1;
2009         }
2010
2011         return 0;
2012 #else
2013         log_err("fio: direct IO is not supported on this host operating system\n");
2014         return -1;
2015 #endif
2016 }