filesetup.c

   1 #include <unistd.h>
   2 #include <fcntl.h>
   3 #include <string.h>
   4 #include <assert.h>
   5 #include <dirent.h>
   6 #include <libgen.h>
   7 #include <sys/stat.h>
   8 #include <sys/mman.h>
   9 #include <sys/types.h>
  10
  11 #include "fio.h"
  12 #include "smalloc.h"
  13 #include "filehash.h"
  14 #include "options.h"
  15 #include "os/os.h"
  16 #include "hash.h"
  17 #include "lib/axmap.h"
  18 #include "rwlock.h"
  19
  20 #ifdef CONFIG_LINUX_FALLOCATE
  21 #include <linux/falloc.h>
  22 #endif
  23
  24 static FLIST_HEAD(filename_list);
  25
  26 /*
  27  * List entry for filename_list
  28  */
  29 struct file_name {
  30         struct flist_head list;
  31         char *filename;
  32 };
  33
  34 static inline void clear_error(struct thread_data *td)
  35 {
  36         td->error = 0;
  37         td->verror[0] = '\0';
  38 }
  39
  40 static int native_fallocate(struct thread_data *td, struct fio_file *f)
  41 {
  42         bool success;
  43
  44         success = fio_fallocate(f, 0, f->real_file_size);
  45         dprint(FD_FILE, "native fallocate of file %s size %llu was "
  46                         "%ssuccessful\n", f->file_name,
  47                         (unsigned long long) f->real_file_size,
  48                         !success ? "un": "");
  49
  50         if (success)
  51                 return false;
  52
  53         if (errno == ENOSYS)
  54                 dprint(FD_FILE, "native fallocate is not implemented\n");
  55
  56         return true;
  57 }
  58
  59 static void fallocate_file(struct thread_data *td, struct fio_file *f)
  60 {
  61         if (td->o.fill_device)
  62                 return;
  63
  64         switch (td->o.fallocate_mode) {
  65         case FIO_FALLOCATE_NATIVE:
  66                 native_fallocate(td, f);
  67                 break;
  68         case FIO_FALLOCATE_NONE:
  69                 break;
  70 #ifdef CONFIG_POSIX_FALLOCATE
  71         case FIO_FALLOCATE_POSIX: {
  72                 int r;
  73
  74                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  75                                  f->file_name,
  76                                  (unsigned long long) f->real_file_size);
  77
  78                 r = posix_fallocate(f->fd, 0, f->real_file_size);
  79                 if (r > 0)
  80                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
  81                 break;
  82                 }
  83 #endif /* CONFIG_POSIX_FALLOCATE */
  84 #ifdef CONFIG_LINUX_FALLOCATE
  85         case FIO_FALLOCATE_KEEP_SIZE: {
  86                 int r;
  87
  88                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
  89                                 "file %s size %llu\n", f->file_name,
  90                                 (unsigned long long) f->real_file_size);
  91
  92                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
  93                 if (r != 0)
  94                         td_verror(td, errno, "fallocate");
  95
  96                 break;
  97                 }
  98 #endif /* CONFIG_LINUX_FALLOCATE */
  99         default:
 100                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
 101                 assert(0);
 102         }
 103 }
 104
 105 /*
 106  * Leaves f->fd open on success, caller must close
 107  */
 108 static int extend_file(struct thread_data *td, struct fio_file *f)
 109 {
 110         int new_layout = 0, unlink_file = 0, flags;
 111         unsigned long long left;
 112         unsigned int bs;
 113         char *b = NULL;
 114
 115         if (read_only) {
 116                 log_err("fio: refusing extend of file due to read-only\n");
 117                 return 0;
 118         }
 119
 120         /*
 121          * check if we need to lay the file out complete again. fio
 122          * does that for operations involving reads, or for writes
 123          * where overwrite is set
 124          */
 125         if (td_read(td) ||
 126            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
 127             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
 128                 new_layout = 1;
 129         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
 130                 unlink_file = 1;
 131
 132         if (unlink_file || new_layout) {
 133                 int ret;
 134
 135                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
 136
 137                 ret = td_io_unlink_file(td, f);
 138                 if (ret != 0 && ret != ENOENT) {
 139                         td_verror(td, errno, "unlink");
 140                         return 1;
 141                 }
 142         }
 143
 144         flags = O_WRONLY;
 145         if (td->o.allow_create)
 146                 flags |= O_CREAT;
 147         if (new_layout)
 148                 flags |= O_TRUNC;
 149
 150 #ifdef WIN32
 151         flags |= _O_BINARY;
 152 #endif
 153
 154         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 155         f->fd = open(f->file_name, flags, 0644);
 156         if (f->fd < 0) {
 157                 int err = errno;
 158
 159                 if (err == ENOENT && !td->o.allow_create)
 160                         log_err("fio: file creation disallowed by "
 161                                         "allow_file_create=0\n");
 162                 else
 163                         td_verror(td, err, "open");
 164                 return 1;
 165         }
 166
 167         fallocate_file(td, f);
 168
 169         /*
 170          * If our jobs don't require regular files initially, we're done.
 171          */
 172         if (!new_layout)
 173                 goto done;
 174
 175         /*
 176          * The size will be -1ULL when fill_device is used, so don't truncate
 177          * or fallocate this file, just write it
 178          */
 179         if (!td->o.fill_device) {
 180                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 181                                         (unsigned long long) f->real_file_size);
 182                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 183                         if (errno != EFBIG) {
 184                                 td_verror(td, errno, "ftruncate");
 185                                 goto err;
 186                         }
 187                 }
 188         }
 189
 190         left = f->real_file_size;
 191         bs = td->o.max_bs[DDIR_WRITE];
 192         if (bs > left)
 193                 bs = left;
 194
 195         b = malloc(bs);
 196         if (!b) {
 197                 td_verror(td, errno, "malloc");
 198                 goto err;
 199         }
 200
 201         while (left && !td->terminate) {
 202                 ssize_t r;
 203
 204                 if (bs > left)
 205                         bs = left;
 206
 207                 fill_io_buffer(td, b, bs, bs);
 208
 209                 r = write(f->fd, b, bs);
 210
 211                 if (r > 0) {
 212                         left -= r;
 213                         continue;
 214                 } else {
 215                         if (r < 0) {
 216                                 int __e = errno;
 217
 218                                 if (__e == ENOSPC) {
 219                                         if (td->o.fill_device)
 220                                                 break;
 221                                         log_info("fio: ENOSPC on laying out "
 222                                                  "file, stopping\n");
 223                                         break;
 224                                 }
 225                                 td_verror(td, errno, "write");
 226                         } else
 227                                 td_verror(td, EIO, "write");
 228
 229                         break;
 230                 }
 231         }
 232
 233         if (td->terminate) {
 234                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 235                 td_io_unlink_file(td, f);
 236         } else if (td->o.create_fsync) {
 237                 if (fsync(f->fd) < 0) {
 238                         td_verror(td, errno, "fsync");
 239                         goto err;
 240                 }
 241         }
 242         if (td->o.fill_device && !td_write(td)) {
 243                 fio_file_clear_size_known(f);
 244                 if (td_io_get_file_size(td, f))
 245                         goto err;
 246                 if (f->io_size > f->real_file_size)
 247                         f->io_size = f->real_file_size;
 248         }
 249
 250         free(b);
 251 done:
 252         return 0;
 253 err:
 254         close(f->fd);
 255         f->fd = -1;
 256         if (b)
 257                 free(b);
 258         return 1;
 259 }
 260
 261 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 262 {
 263         int r, did_open = 0, old_runstate;
 264         unsigned long long left;
 265         unsigned int bs;
 266         bool ret = true;
 267         char *b;
 268
 269         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 270             td_ioengine_flagged(td, FIO_NOIO))
 271                 return true;
 272
 273         if (f->filetype == FIO_TYPE_CHAR)
 274                 return true;
 275
 276         if (!fio_file_open(f)) {
 277                 if (td->io_ops->open_file(td, f)) {
 278                         log_err("fio: cannot pre-read, failed to open file\n");
 279                         return false;
 280                 }
 281                 did_open = 1;
 282         }
 283
 284         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 285
 286         left = f->io_size;
 287         bs = td->o.max_bs[DDIR_READ];
 288         if (bs > left)
 289                 bs = left;
 290
 291         b = malloc(bs);
 292         if (!b) {
 293                 td_verror(td, errno, "malloc");
 294                 ret = false;
 295                 goto error;
 296         }
 297         memset(b, 0, bs);
 298
 299         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 300                 td_verror(td, errno, "lseek");
 301                 log_err("fio: failed to lseek pre-read file\n");
 302                 ret = false;
 303                 goto error;
 304         }
 305
 306         while (left && !td->terminate) {
 307                 if (bs > left)
 308                         bs = left;
 309
 310                 r = read(f->fd, b, bs);
 311
 312                 if (r == (int) bs) {
 313                         left -= bs;
 314                         continue;
 315                 } else {
 316                         td_verror(td, EIO, "pre_read");
 317                         break;
 318                 }
 319         }
 320
 321 error:
 322         td_restore_runstate(td, old_runstate);
 323
 324         if (did_open)
 325                 td->io_ops->close_file(td, f);
 326
 327         free(b);
 328         return ret;
 329 }
 330
 331 unsigned long long get_rand_file_size(struct thread_data *td)
 332 {
 333         unsigned long long ret, sized;
 334         uint64_t frand_max;
 335         unsigned long r;
 336
 337         frand_max = rand_max(&td->file_size_state);
 338         r = __rand(&td->file_size_state);
 339         sized = td->o.file_size_high - td->o.file_size_low;
 340         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 341         ret += td->o.file_size_low;
 342         ret -= (ret % td->o.rw_min_bs);
 343         return ret;
 344 }
 345
 346 static int file_size(struct thread_data *td, struct fio_file *f)
 347 {
 348         struct stat st;
 349
 350         if (stat(f->file_name, &st) == -1) {
 351                 td_verror(td, errno, "fstat");
 352                 return 1;
 353         }
 354
 355         f->real_file_size = st.st_size;
 356         return 0;
 357 }
 358
 359 static int bdev_size(struct thread_data *td, struct fio_file *f)
 360 {
 361         unsigned long long bytes = 0;
 362         int r;
 363
 364         if (td->io_ops->open_file(td, f)) {
 365                 log_err("fio: failed opening blockdev %s for size check\n",
 366                         f->file_name);
 367                 return 1;
 368         }
 369
 370         r = blockdev_size(f, &bytes);
 371         if (r) {
 372                 td_verror(td, r, "blockdev_size");
 373                 goto err;
 374         }
 375
 376         if (!bytes) {
 377                 log_err("%s: zero sized block device?\n", f->file_name);
 378                 goto err;
 379         }
 380
 381         f->real_file_size = bytes;
 382         td->io_ops->close_file(td, f);
 383         return 0;
 384 err:
 385         td->io_ops->close_file(td, f);
 386         return 1;
 387 }
 388
 389 static int char_size(struct thread_data *td, struct fio_file *f)
 390 {
 391 #ifdef FIO_HAVE_CHARDEV_SIZE
 392         unsigned long long bytes = 0;
 393         int r;
 394
 395         if (td->io_ops->open_file(td, f)) {
 396                 log_err("fio: failed opening chardev %s for size check\n",
 397                         f->file_name);
 398                 return 1;
 399         }
 400
 401         r = chardev_size(f, &bytes);
 402         if (r) {
 403                 td_verror(td, r, "chardev_size");
 404                 goto err;
 405         }
 406
 407         if (!bytes) {
 408                 log_err("%s: zero sized char device?\n", f->file_name);
 409                 goto err;
 410         }
 411
 412         f->real_file_size = bytes;
 413         td->io_ops->close_file(td, f);
 414         return 0;
 415 err:
 416         td->io_ops->close_file(td, f);
 417         return 1;
 418 #else
 419         f->real_file_size = -1ULL;
 420         return 0;
 421 #endif
 422 }
 423
 424 static int get_file_size(struct thread_data *td, struct fio_file *f)
 425 {
 426         int ret = 0;
 427
 428         if (fio_file_size_known(f))
 429                 return 0;
 430
 431         if (f->filetype == FIO_TYPE_FILE)
 432                 ret = file_size(td, f);
 433         else if (f->filetype == FIO_TYPE_BLOCK)
 434                 ret = bdev_size(td, f);
 435         else if (f->filetype == FIO_TYPE_CHAR)
 436                 ret = char_size(td, f);
 437         else {
 438                 f->real_file_size = -1;
 439                 log_info("%s: failed to get file size of %s\n", td->o.name,
 440                                         f->file_name);
 441                 return 1; /* avoid offset extends end error message */
 442         }
 443
 444         /*
 445          * Leave ->real_file_size with 0 since it could be expectation
 446          * of initial setup for regular files.
 447          */
 448         if (ret)
 449                 return ret;
 450
 451         /*
 452          * ->file_offset normally hasn't been initialized yet, so this
 453          * is basically always false unless ->real_file_size is -1, but
 454          * if ->real_file_size is -1 this message doesn't make sense.
 455          * As a result, this message is basically useless.
 456          */
 457         if (f->file_offset > f->real_file_size) {
 458                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 459                                         (unsigned long long) f->file_offset,
 460                                         (unsigned long long) f->real_file_size);
 461                 return 1;
 462         }
 463
 464         fio_file_set_size_known(f);
 465         return 0;
 466 }
 467
 468 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 469                                    unsigned long long off,
 470                                    unsigned long long len)
 471 {
 472         int errval = 0, ret = 0;
 473
 474 #ifdef CONFIG_ESX
 475         return 0;
 476 #endif
 477
 478         if (len == -1ULL)
 479                 len = f->io_size;
 480         if (off == -1ULL)
 481                 off = f->file_offset;
 482
 483         if (len == -1ULL || off == -1ULL)
 484                 return 0;
 485
 486         if (td->io_ops->invalidate) {
 487                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 488                         f->file_name);
 489                 ret = td->io_ops->invalidate(td, f);
 490                 if (ret < 0)
 491                         errval = -ret;
 492         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 493                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
 494                        td->io_ops->name);
 495                 ret = 0;
 496         } else if (f->filetype == FIO_TYPE_FILE) {
 497                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 498                         f->file_name, off, len);
 499                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 500                 if (ret)
 501                         errval = ret;
 502         } else if (f->filetype == FIO_TYPE_BLOCK) {
 503                 int retry_count = 0;
 504
 505                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 506                 ret = blockdev_invalidate_cache(f);
 507                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 508                         /*
 509                          * Linux multipath devices reject ioctl while
 510                          * the maps are being updated. That window can
 511                          * last tens of milliseconds; we'll try up to
 512                          * a quarter of a second.
 513                          */
 514                         usleep(10000);
 515                         ret = blockdev_invalidate_cache(f);
 516                 }
 517                 if (ret < 0 && errno == EACCES && geteuid()) {
 518                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 519                                 log_err("fio: only root may flush block "
 520                                         "devices. Cache flush bypassed!\n");
 521                         }
 522                         ret = 0;
 523                 }
 524                 if (ret < 0)
 525                         errval = errno;
 526         } else if (f->filetype == FIO_TYPE_CHAR ||
 527                    f->filetype == FIO_TYPE_PIPE) {
 528                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 529                 ret = 0;
 530         }
 531
 532         /*
 533          * Cache flushing isn't a fatal condition, and we know it will
 534          * happen on some platforms where we don't have the proper
 535          * function to flush eg block device caches. So just warn and
 536          * continue on our way.
 537          */
 538         if (errval)
 539                 log_info("fio: cache invalidation of %s failed: %s\n",
 540                          f->file_name, strerror(errval));
 541
 542         return 0;
 543
 544 }
 545
 546 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 547 {
 548         if (!fio_file_open(f))
 549                 return 0;
 550
 551         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 552 }
 553
 554 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 555 {
 556         int ret = 0;
 557
 558         dprint(FD_FILE, "fd close %s\n", f->file_name);
 559
 560         remove_file_hash(f);
 561
 562         if (close(f->fd) < 0)
 563                 ret = errno;
 564
 565         f->fd = -1;
 566
 567         if (f->shadow_fd != -1) {
 568                 close(f->shadow_fd);
 569                 f->shadow_fd = -1;
 570         }
 571
 572         f->engine_pos = 0;
 573         return ret;
 574 }
 575
 576 int file_lookup_open(struct fio_file *f, int flags)
 577 {
 578         struct fio_file *__f;
 579         int from_hash;
 580
 581         __f = lookup_file_hash(f->file_name);
 582         if (__f) {
 583                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 584                 f->lock = __f->lock;
 585                 from_hash = 1;
 586         } else {
 587                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 588                 from_hash = 0;
 589         }
 590
 591 #ifdef WIN32
 592         flags |= _O_BINARY;
 593 #endif
 594
 595         f->fd = open(f->file_name, flags, 0600);
 596         return from_hash;
 597 }
 598
 599 static int file_close_shadow_fds(struct thread_data *td)
 600 {
 601         struct fio_file *f;
 602         int num_closed = 0;
 603         unsigned int i;
 604
 605         for_each_file(td, f, i) {
 606                 if (f->shadow_fd == -1)
 607                         continue;
 608
 609                 close(f->shadow_fd);
 610                 f->shadow_fd = -1;
 611                 num_closed++;
 612         }
 613
 614         return num_closed;
 615 }
 616
 617 int generic_open_file(struct thread_data *td, struct fio_file *f)
 618 {
 619         int is_std = 0;
 620         int flags = 0;
 621         int from_hash = 0;
 622
 623         dprint(FD_FILE, "fd open %s\n", f->file_name);
 624
 625         if (!strcmp(f->file_name, "-")) {
 626                 if (td_rw(td)) {
 627                         log_err("fio: can't read/write to stdin/out\n");
 628                         return 1;
 629                 }
 630                 is_std = 1;
 631
 632                 /*
 633                  * move output logging to stderr, if we are writing to stdout
 634                  */
 635                 if (td_write(td))
 636                         f_out = stderr;
 637         }
 638
 639         if (td_trim(td))
 640                 goto skip_flags;
 641         if (td->o.odirect)
 642                 flags |= OS_O_DIRECT;
 643         if (td->o.oatomic) {
 644                 if (!FIO_O_ATOMIC) {
 645                         td_verror(td, EINVAL, "OS does not support atomic IO");
 646                         return 1;
 647                 }
 648                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 649         }
 650         if (td->o.sync_io)
 651                 flags |= O_SYNC;
 652         if (td->o.create_on_open && td->o.allow_create)
 653                 flags |= O_CREAT;
 654 skip_flags:
 655         if (f->filetype != FIO_TYPE_FILE)
 656                 flags |= FIO_O_NOATIME;
 657
 658 open_again:
 659         if (td_write(td)) {
 660                 if (!read_only)
 661                         flags |= O_RDWR;
 662
 663                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 664                         flags |= O_CREAT;
 665
 666                 if (is_std)
 667                         f->fd = dup(STDOUT_FILENO);
 668                 else
 669                         from_hash = file_lookup_open(f, flags);
 670         } else if (td_read(td)) {
 671                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 672                         flags |= O_RDWR;
 673                 else
 674                         flags |= O_RDONLY;
 675
 676                 if (is_std)
 677                         f->fd = dup(STDIN_FILENO);
 678                 else
 679                         from_hash = file_lookup_open(f, flags);
 680         } else if (td_trim(td)) {
 681                 assert(!td_rw(td)); /* should have matched above */
 682                 flags |= O_RDWR;
 683                 from_hash = file_lookup_open(f, flags);
 684         }
 685
 686         if (f->fd == -1) {
 687                 char buf[FIO_VERROR_SIZE];
 688                 int __e = errno;
 689
 690                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 691                         flags &= ~FIO_O_NOATIME;
 692                         goto open_again;
 693                 }
 694                 if (__e == EMFILE && file_close_shadow_fds(td))
 695                         goto open_again;
 696
 697                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 698
 699                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 700                         log_err("fio: looks like your file system does not " \
 701                                 "support direct=1/buffered=0\n");
 702                 }
 703
 704                 td_verror(td, __e, buf);
 705                 return 1;
 706         }
 707
 708         if (!from_hash && f->fd != -1) {
 709                 if (add_file_hash(f)) {
 710                         int fio_unused ret;
 711
 712                         /*
 713                          * Stash away descriptor for later close. This is to
 714                          * work-around a "feature" on Linux, where a close of
 715                          * an fd that has been opened for write will trigger
 716                          * udev to call blkid to check partitions, fs id, etc.
 717                          * That pollutes the device cache, which can slow down
 718                          * unbuffered accesses.
 719                          */
 720                         if (f->shadow_fd == -1)
 721                                 f->shadow_fd = f->fd;
 722                         else {
 723                                 /*
 724                                  * OK to ignore, we haven't done anything
 725                                  * with it
 726                                  */
 727                                 ret = generic_close_file(td, f);
 728                         }
 729                         goto open_again;
 730                 }
 731         }
 732
 733         return 0;
 734 }
 735
 736 /*
 737  * This function i.e. get_file_size() is the default .get_file_size
 738  * implementation of majority of I/O engines.
 739  */
 740 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 741 {
 742         return get_file_size(td, f);
 743 }
 744
 745 /*
 746  * open/close all files, so that ->real_file_size gets set
 747  */
 748 static int get_file_sizes(struct thread_data *td)
 749 {
 750         struct fio_file *f;
 751         unsigned int i;
 752         int err = 0;
 753
 754         for_each_file(td, f, i) {
 755                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 756                                                                 f->file_name);
 757
 758                 if (td_io_get_file_size(td, f)) {
 759                         if (td->error != ENOENT) {
 760                                 log_err("%s\n", td->verror);
 761                                 err = 1;
 762                                 break;
 763                         }
 764                         clear_error(td);
 765                 }
 766
 767                 /*
 768                  * There are corner cases where we end up with -1 for
 769                  * ->real_file_size due to unsupported file type, etc.
 770                  * We then just set to size option value divided by number
 771                  * of files, similar to the way file ->io_size is set.
 772                  * stat(2) failure doesn't set ->real_file_size to -1.
 773                  */
 774                 if (f->real_file_size == -1ULL && td->o.size)
 775                         f->real_file_size = td->o.size / td->o.nr_files;
 776         }
 777
 778         return err;
 779 }
 780
 781 struct fio_mount {
 782         struct flist_head list;
 783         const char *base;
 784         char __base[256];
 785         unsigned int key;
 786 };
 787
 788 /*
 789  * Get free number of bytes for each file on each unique mount.
 790  */
 791 static unsigned long long get_fs_free_counts(struct thread_data *td)
 792 {
 793         struct flist_head *n, *tmp;
 794         unsigned long long ret = 0;
 795         struct fio_mount *fm;
 796         FLIST_HEAD(list);
 797         struct fio_file *f;
 798         unsigned int i;
 799
 800         for_each_file(td, f, i) {
 801                 struct stat sb;
 802                 char buf[256];
 803
 804                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 805                         if (f->real_file_size != -1ULL)
 806                                 ret += f->real_file_size;
 807                         continue;
 808                 } else if (f->filetype != FIO_TYPE_FILE)
 809                         continue;
 810
 811                 buf[255] = '\0';
 812                 strncpy(buf, f->file_name, 255);
 813
 814                 if (stat(buf, &sb) < 0) {
 815                         if (errno != ENOENT)
 816                                 break;
 817                         strcpy(buf, ".");
 818                         if (stat(buf, &sb) < 0)
 819                                 break;
 820                 }
 821
 822                 fm = NULL;
 823                 flist_for_each(n, &list) {
 824                         fm = flist_entry(n, struct fio_mount, list);
 825                         if (fm->key == sb.st_dev)
 826                                 break;
 827
 828                         fm = NULL;
 829                 }
 830
 831                 if (fm)
 832                         continue;
 833
 834                 fm = calloc(1, sizeof(*fm));
 835                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 836                 fm->base = basename(fm->__base);
 837                 fm->key = sb.st_dev;
 838                 flist_add(&fm->list, &list);
 839         }
 840
 841         flist_for_each_safe(n, tmp, &list) {
 842                 unsigned long long sz;
 843
 844                 fm = flist_entry(n, struct fio_mount, list);
 845                 flist_del(&fm->list);
 846
 847                 sz = get_fs_free_size(fm->base);
 848                 if (sz && sz != -1ULL)
 849                         ret += sz;
 850
 851                 free(fm);
 852         }
 853
 854         return ret;
 855 }
 856
 857 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 858 {
 859         struct thread_options *o = &td->o;
 860         unsigned long long align_bs;
 861         unsigned long long offset;
 862
 863         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 864                 return f->real_file_size;
 865
 866         if (o->start_offset_percent > 0) {
 867                 /*
 868                  * if offset_align is provided, set initial offset
 869                  */
 870                 if (fio_option_is_set(o, start_offset_align)) {
 871                         align_bs = o->start_offset_align;
 872                 } else {
 873                         /* else take the minimum block size */
 874                         align_bs = td_min_bs(td);
 875                 }
 876
 877                 /* calculate the raw offset */
 878                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 879                         (td->subjob_number * o->offset_increment);
 880
 881                 /*
 882                  * block align the offset at the next available boundary at
 883                  * ceiling(offset / align_bs) * align_bs
 884                  */
 885                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 886
 887         } else {
 888                 /* start_offset_percent not set */
 889                 offset = o->start_offset +
 890                                 td->subjob_number * o->offset_increment;
 891         }
 892
 893         return offset;
 894 }
 895
 896 /*
 897  * Open the files and setup files sizes, creating files if necessary.
 898  */
 899 int setup_files(struct thread_data *td)
 900 {
 901         unsigned long long total_size, extend_size;
 902         struct thread_options *o = &td->o;
 903         struct fio_file *f;
 904         unsigned int i, nr_fs_extra = 0;
 905         int err = 0, need_extend;
 906         int old_state;
 907         const unsigned int bs = td_min_bs(td);
 908         uint64_t fs = 0;
 909
 910         dprint(FD_FILE, "setup files\n");
 911
 912         old_state = td_bump_runstate(td, TD_SETTING_UP);
 913
 914         if (o->read_iolog_file)
 915                 goto done;
 916
 917         /*
 918          * Find out physical size of files or devices for this thread,
 919          * before we determine I/O size and range of our targets.
 920          * If ioengine defines a setup() method, it's responsible for
 921          * opening the files and setting f->real_file_size to indicate
 922          * the valid range for that file.
 923          */
 924         if (td->io_ops->setup)
 925                 err = td->io_ops->setup(td);
 926         else
 927                 err = get_file_sizes(td);
 928
 929         if (err)
 930                 goto err_out;
 931
 932         /*
 933          * check sizes. if the files/devices do not exist and the size
 934          * isn't passed to fio, abort.
 935          */
 936         total_size = 0;
 937         for_each_file(td, f, i) {
 938                 f->fileno = i;
 939                 if (f->real_file_size == -1ULL)
 940                         total_size = -1ULL;
 941                 else
 942                         total_size += f->real_file_size;
 943         }
 944
 945         if (o->fill_device)
 946                 td->fill_device_size = get_fs_free_counts(td);
 947
 948         /*
 949          * device/file sizes are zero and no size given, punt
 950          */
 951         if ((!total_size || total_size == -1ULL) && !o->size &&
 952             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 953             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 954                 log_err("%s: you need to specify size=\n", o->name);
 955                 td_verror(td, EINVAL, "total_file_size");
 956                 goto err_out;
 957         }
 958
 959         /*
 960          * Calculate per-file size and potential extra size for the
 961          * first files, if needed (i.e. if we don't have a fixed size).
 962          */
 963         if (!o->file_size_low && o->nr_files) {
 964                 uint64_t all_fs;
 965
 966                 fs = o->size / o->nr_files;
 967                 all_fs = fs * o->nr_files;
 968
 969                 if (all_fs < o->size)
 970                         nr_fs_extra = (o->size - all_fs) / bs;
 971         }
 972
 973         /*
 974          * now file sizes are known, so we can set ->io_size. if size= is
 975          * not given, ->io_size is just equal to ->real_file_size. if size
 976          * is given, ->io_size is size / nr_files.
 977          */
 978         extend_size = total_size = 0;
 979         need_extend = 0;
 980         for_each_file(td, f, i) {
 981                 f->file_offset = get_start_offset(td, f);
 982
 983                 /*
 984                  * Update ->io_size depending on options specified.
 985                  * ->file_size_low being 0 means filesize option isn't set.
 986                  * Non zero ->file_size_low equals ->file_size_high means
 987                  * filesize option is set in a fixed size format.
 988                  * Non zero ->file_size_low not equals ->file_size_high means
 989                  * filesize option is set in a range format.
 990                  */
 991                 if (!o->file_size_low) {
 992                         /*
 993                          * no file size or range given, file size is equal to
 994                          * total size divided by number of files. If the size
 995                          * doesn't divide nicely with the min blocksize,
 996                          * make the first files bigger.
 997                          */
 998                         f->io_size = fs;
 999                         if (nr_fs_extra) {
1000                                 nr_fs_extra--;
1001                                 f->io_size += bs;
1002                         }
1003
1004                         /*
1005                          * We normally don't come here for regular files, but
1006                          * if the result is 0 for a regular file, set it to the
1007                          * real file size. This could be size of the existing
1008                          * one if it already exists, but otherwise will be set
1009                          * to 0. A new file won't be created because
1010                          * ->io_size + ->file_offset equals ->real_file_size.
1011                          */
1012                         if (!f->io_size) {
1013                                 if (f->file_offset > f->real_file_size)
1014                                         goto err_offset;
1015                                 f->io_size = f->real_file_size - f->file_offset;
1016                                 if (!f->io_size)
1017                                         log_info("fio: file %s may be ignored\n",
1018                                                 f->file_name);
1019                         }
1020                 } else if (f->real_file_size < o->file_size_low ||
1021                            f->real_file_size > o->file_size_high) {
1022                         if (f->file_offset > o->file_size_low)
1023                                 goto err_offset;
1024                         /*
1025                          * file size given. if it's fixed, use that. if it's a
1026                          * range, generate a random size in-between.
1027                          */
1028                         if (o->file_size_low == o->file_size_high)
1029                                 f->io_size = o->file_size_low - f->file_offset;
1030                         else {
1031                                 f->io_size = get_rand_file_size(td)
1032                                                 - f->file_offset;
1033                         }
1034                 } else
1035                         f->io_size = f->real_file_size - f->file_offset;
1036
1037                 if (f->io_size == -1ULL)
1038                         total_size = -1ULL;
1039                 else {
1040                         if (o->size_percent && o->size_percent != 100) {
1041                                 uint64_t file_size;
1042
1043                                 file_size = f->io_size + f->file_offset;
1044                                 f->io_size = (file_size *
1045                                               o->size_percent) / 100;
1046                                 if (f->io_size > (file_size - f->file_offset))
1047                                         f->io_size = file_size - f->file_offset;
1048
1049                                 f->io_size -= (f->io_size % td_min_bs(td));
1050                         }
1051                         total_size += f->io_size;
1052                 }
1053
1054                 if (f->filetype == FIO_TYPE_FILE &&
1055                     (f->io_size + f->file_offset) > f->real_file_size &&
1056                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1057                         if (!o->create_on_open) {
1058                                 need_extend++;
1059                                 extend_size += (f->io_size + f->file_offset);
1060                                 fio_file_set_extend(f);
1061                         } else
1062                                 f->real_file_size = f->io_size + f->file_offset;
1063                 }
1064         }
1065
1066         if (td->o.block_error_hist) {
1067                 int len;
1068
1069                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1070                 f = td->files[0];
1071                 len = f->io_size / td->o.bs[DDIR_TRIM];
1072                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1073                         log_err("fio: cannot calculate block histogram with "
1074                                 "%d trim blocks, maximum %d\n",
1075                                 len, MAX_NR_BLOCK_INFOS);
1076                         td_verror(td, EINVAL, "block_error_hist");
1077                         goto err_out;
1078                 }
1079
1080                 td->ts.nr_block_infos = len;
1081                 for (i = 0; i < len; i++)
1082                         td->ts.block_infos[i] =
1083                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1084         } else
1085                 td->ts.nr_block_infos = 0;
1086
1087         if (!o->size || (total_size && o->size > total_size))
1088                 o->size = total_size;
1089
1090         if (o->size < td_min_bs(td)) {
1091                 log_err("fio: blocksize too large for data set\n");
1092                 goto err_out;
1093         }
1094
1095         /*
1096          * See if we need to extend some files, typically needed when our
1097          * target regular files don't exist yet, but our jobs require them
1098          * initially due to read I/Os.
1099          */
1100         if (need_extend) {
1101                 temp_stall_ts = 1;
1102                 if (output_format & FIO_OUTPUT_NORMAL) {
1103                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1104                                  o->name,
1105                                  need_extend > 1 ? "s" : "",
1106                                  need_extend,
1107                                  need_extend > 1 ? "s" : "",
1108                                  need_extend > 1 ? "total " : "",
1109                                  extend_size >> 20);
1110                 }
1111
1112                 for_each_file(td, f, i) {
1113                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1114
1115                         if (!fio_file_extend(f))
1116                                 continue;
1117
1118                         assert(f->filetype == FIO_TYPE_FILE);
1119                         fio_file_clear_extend(f);
1120                         if (!o->fill_device) {
1121                                 old_len = f->real_file_size;
1122                                 extend_len = f->io_size + f->file_offset -
1123                                                 old_len;
1124                         }
1125                         f->real_file_size = (f->io_size + f->file_offset);
1126                         err = extend_file(td, f);
1127                         if (err)
1128                                 break;
1129
1130                         err = __file_invalidate_cache(td, f, old_len,
1131                                                                 extend_len);
1132
1133                         /*
1134                          * Shut up static checker
1135                          */
1136                         if (f->fd != -1)
1137                                 close(f->fd);
1138
1139                         f->fd = -1;
1140                         if (err)
1141                                 break;
1142                 }
1143                 temp_stall_ts = 0;
1144         }
1145
1146         if (err)
1147                 goto err_out;
1148
1149         if (!o->zone_size)
1150                 o->zone_size = o->size;
1151
1152         /*
1153          * iolog already set the total io size, if we read back
1154          * stored entries.
1155          */
1156         if (!o->read_iolog_file) {
1157                 if (o->io_size)
1158                         td->total_io_size = o->io_size * o->loops;
1159                 else
1160                         td->total_io_size = o->size * o->loops;
1161         }
1162
1163 done:
1164         if (o->create_only)
1165                 td->done = 1;
1166
1167         td_restore_runstate(td, old_state);
1168         return 0;
1169 err_offset:
1170         log_err("%s: you need to specify valid offset=\n", o->name);
1171 err_out:
1172         td_restore_runstate(td, old_state);
1173         return 1;
1174 }
1175
1176 bool pre_read_files(struct thread_data *td)
1177 {
1178         struct fio_file *f;
1179         unsigned int i;
1180
1181         dprint(FD_FILE, "pre_read files\n");
1182
1183         for_each_file(td, f, i) {
1184                 if (!pre_read_file(td, f))
1185                         return false;
1186         }
1187
1188         return true;
1189 }
1190
1191 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1192 {
1193         unsigned int range_size, seed;
1194         unsigned long nranges;
1195         uint64_t fsize;
1196
1197         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1198         fsize = min(f->real_file_size, f->io_size);
1199
1200         nranges = (fsize + range_size - 1) / range_size;
1201
1202         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1203         if (!td->o.rand_repeatable)
1204                 seed = td->rand_seeds[4];
1205
1206         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1207                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1208         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1209                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1210         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1211                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1212 }
1213
1214 static bool init_rand_distribution(struct thread_data *td)
1215 {
1216         struct fio_file *f;
1217         unsigned int i;
1218         int state;
1219
1220         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1221                 return false;
1222
1223         state = td_bump_runstate(td, TD_SETTING_UP);
1224
1225         for_each_file(td, f, i)
1226                 __init_rand_distribution(td, f);
1227
1228         td_restore_runstate(td, state);
1229         return true;
1230 }
1231
1232 /*
1233  * Check if the number of blocks exceeds the randomness capability of
1234  * the selected generator. Tausworthe is 32-bit, the others are fullly
1235  * 64-bit capable.
1236  */
1237 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1238                                  uint64_t blocks)
1239 {
1240         if (blocks <= FRAND32_MAX)
1241                 return 0;
1242         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1243                 return 0;
1244
1245         /*
1246          * If the user hasn't specified a random generator, switch
1247          * to tausworthe64 with informational warning. If the user did
1248          * specify one, just warn.
1249          */
1250         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1251                         f->file_name);
1252
1253         if (!fio_option_is_set(&td->o, random_generator)) {
1254                 log_info("fio: Switching to tausworthe64. Use the "
1255                          "random_generator= option to get rid of this "
1256                          "warning.\n");
1257                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1258                 return 0;
1259         }
1260
1261         /*
1262          * Just make this information to avoid breaking scripts.
1263          */
1264         log_info("fio: Use the random_generator= option to switch to lfsr or "
1265                          "tausworthe64.\n");
1266         return 0;
1267 }
1268
1269 bool init_random_map(struct thread_data *td)
1270 {
1271         unsigned long long blocks;
1272         struct fio_file *f;
1273         unsigned int i;
1274
1275         if (init_rand_distribution(td))
1276                 return true;
1277         if (!td_random(td))
1278                 return true;
1279
1280         for_each_file(td, f, i) {
1281                 uint64_t fsize = min(f->real_file_size, f->io_size);
1282
1283                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1284
1285                 if (check_rand_gen_limits(td, f, blocks))
1286                         return false;
1287
1288                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1289                         unsigned long seed;
1290
1291                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1292
1293                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1294                                 fio_file_set_lfsr(f);
1295                                 continue;
1296                         }
1297                 } else if (!td->o.norandommap) {
1298                         f->io_axmap = axmap_new(blocks);
1299                         if (f->io_axmap) {
1300                                 fio_file_set_axmap(f);
1301                                 continue;
1302                         }
1303                 } else if (td->o.norandommap)
1304                         continue;
1305
1306                 if (!td->o.softrandommap) {
1307                         log_err("fio: failed allocating random map. If running"
1308                                 " a large number of jobs, try the 'norandommap'"
1309                                 " option or set 'softrandommap'. Or give"
1310                                 " a larger --alloc-size to fio.\n");
1311                         return false;
1312                 }
1313
1314                 log_info("fio: file %s failed allocating random map. Running "
1315                          "job without.\n", f->file_name);
1316         }
1317
1318         return true;
1319 }
1320
1321 void close_files(struct thread_data *td)
1322 {
1323         struct fio_file *f;
1324         unsigned int i;
1325
1326         for_each_file(td, f, i) {
1327                 if (fio_file_open(f))
1328                         td_io_close_file(td, f);
1329         }
1330 }
1331
1332 void close_and_free_files(struct thread_data *td)
1333 {
1334         struct fio_file *f;
1335         unsigned int i;
1336         bool use_free = td_ioengine_flagged(td, FIO_NOFILEHASH);
1337
1338         dprint(FD_FILE, "close files\n");
1339
1340         for_each_file(td, f, i) {
1341                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1342                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1343                         td_io_unlink_file(td, f);
1344                 }
1345
1346                 if (fio_file_open(f))
1347                         td_io_close_file(td, f);
1348
1349                 remove_file_hash(f);
1350
1351                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1352                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1353                         td_io_unlink_file(td, f);
1354                 }
1355
1356                 if (use_free)
1357                         free(f->file_name);
1358                 else
1359                         sfree(f->file_name);
1360                 f->file_name = NULL;
1361                 if (fio_file_axmap(f)) {
1362                         axmap_free(f->io_axmap);
1363                         f->io_axmap = NULL;
1364                 }
1365                 if (use_free)
1366                         free(f);
1367                 else
1368                         sfree(f);
1369         }
1370
1371         td->o.filename = NULL;
1372         free(td->files);
1373         free(td->file_locks);
1374         td->files_index = 0;
1375         td->files = NULL;
1376         td->file_locks = NULL;
1377         td->o.file_lock_mode = FILE_LOCK_NONE;
1378         td->o.nr_files = 0;
1379 }
1380
1381 static void get_file_type(struct fio_file *f)
1382 {
1383         struct stat sb;
1384
1385         if (!strcmp(f->file_name, "-"))
1386                 f->filetype = FIO_TYPE_PIPE;
1387         else
1388                 f->filetype = FIO_TYPE_FILE;
1389
1390 #ifdef WIN32
1391         /* \\.\ is the device namespace in Windows, where every file is
1392          * a block device */
1393         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1394                 f->filetype = FIO_TYPE_BLOCK;
1395 #endif
1396
1397         if (!stat(f->file_name, &sb)) {
1398                 if (S_ISBLK(sb.st_mode))
1399                         f->filetype = FIO_TYPE_BLOCK;
1400                 else if (S_ISCHR(sb.st_mode))
1401                         f->filetype = FIO_TYPE_CHAR;
1402                 else if (S_ISFIFO(sb.st_mode))
1403                         f->filetype = FIO_TYPE_PIPE;
1404         }
1405 }
1406
1407 static bool __is_already_allocated(const char *fname, bool set)
1408 {
1409         struct flist_head *entry;
1410         bool ret;
1411
1412         ret = file_bloom_exists(fname, set);
1413         if (!ret)
1414                 return ret;
1415
1416         flist_for_each(entry, &filename_list) {
1417                 struct file_name *fn;
1418
1419                 fn = flist_entry(entry, struct file_name, list);
1420
1421                 if (!strcmp(fn->filename, fname))
1422                         return true;
1423         }
1424
1425         return false;
1426 }
1427
1428 static bool is_already_allocated(const char *fname)
1429 {
1430         bool ret;
1431
1432         fio_file_hash_lock();
1433         ret = __is_already_allocated(fname, false);
1434         fio_file_hash_unlock();
1435
1436         return ret;
1437 }
1438
1439 static void set_already_allocated(const char *fname)
1440 {
1441         struct file_name *fn;
1442
1443         fn = malloc(sizeof(struct file_name));
1444         fn->filename = strdup(fname);
1445
1446         fio_file_hash_lock();
1447         if (!__is_already_allocated(fname, true)) {
1448                 flist_add_tail(&fn->list, &filename_list);
1449                 fn = NULL;
1450         }
1451         fio_file_hash_unlock();
1452
1453         if (fn) {
1454                 free(fn->filename);
1455                 free(fn);
1456         }
1457 }
1458
1459 static void free_already_allocated(void)
1460 {
1461         struct flist_head *entry, *tmp;
1462         struct file_name *fn;
1463
1464         if (flist_empty(&filename_list))
1465                 return;
1466
1467         fio_file_hash_lock();
1468         flist_for_each_safe(entry, tmp, &filename_list) {
1469                 fn = flist_entry(entry, struct file_name, list);
1470                 free(fn->filename);
1471                 flist_del(&fn->list);
1472                 free(fn);
1473         }
1474
1475         fio_file_hash_unlock();
1476 }
1477
1478 static struct fio_file *alloc_new_file(struct thread_data *td)
1479 {
1480         struct fio_file *f;
1481
1482         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1483                 f = calloc(1, sizeof(*f));
1484         else
1485                 f = scalloc(1, sizeof(*f));
1486         if (!f) {
1487                 assert(0);
1488                 return NULL;
1489         }
1490
1491         f->fd = -1;
1492         f->shadow_fd = -1;
1493         fio_file_reset(td, f);
1494         return f;
1495 }
1496
1497 bool exists_and_not_regfile(const char *filename)
1498 {
1499         struct stat sb;
1500
1501         if (lstat(filename, &sb) == -1)
1502                 return false;
1503
1504 #ifndef WIN32 /* NOT Windows */
1505         if (S_ISREG(sb.st_mode))
1506                 return false;
1507 #else
1508         /* \\.\ is the device namespace in Windows, where every file
1509          * is a device node */
1510         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1511                 return false;
1512 #endif
1513
1514         return true;
1515 }
1516
1517 static bool create_work_dirs(struct thread_data *td, const char *fname)
1518 {
1519         char path[PATH_MAX];
1520         char *start, *end;
1521
1522         if (td->o.directory) {
1523                 snprintf(path, PATH_MAX, "%s%c%s", td->o.directory,
1524                          FIO_OS_PATH_SEPARATOR, fname);
1525                 start = strstr(path, fname);
1526         } else {
1527                 snprintf(path, PATH_MAX, "%s", fname);
1528                 start = path;
1529         }
1530
1531         end = start;
1532         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1533                 if (end == start)
1534                         break;
1535                 *end = '\0';
1536                 errno = 0;
1537 #ifdef CONFIG_HAVE_MKDIR_TWO
1538                 if (mkdir(path, 0600) && errno != EEXIST) {
1539 #else
1540                 if (mkdir(path) && errno != EEXIST) {
1541 #endif
1542                         log_err("fio: failed to create dir (%s): %d\n",
1543                                 start, errno);
1544                         return false;
1545                 }
1546                 *end = FIO_OS_PATH_SEPARATOR;
1547                 end++;
1548         }
1549         td->flags |= TD_F_DIRS_CREATED;
1550         return true;
1551 }
1552
1553 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1554 {
1555         int cur_files = td->files_index;
1556         char file_name[PATH_MAX];
1557         struct fio_file *f;
1558         int len = 0;
1559
1560         dprint(FD_FILE, "add file %s\n", fname);
1561
1562         if (td->o.directory)
1563                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1564                                         td->o.unique_filename);
1565
1566         sprintf(file_name + len, "%s", fname);
1567
1568         if (strchr(fname, FIO_OS_PATH_SEPARATOR) &&
1569             !(td->flags & TD_F_DIRS_CREATED) &&
1570             !create_work_dirs(td, fname))
1571                 return 1;
1572
1573         /* clean cloned siblings using existing files */
1574         if (numjob && is_already_allocated(file_name) &&
1575             !exists_and_not_regfile(fname))
1576                 return 0;
1577
1578         f = alloc_new_file(td);
1579
1580         if (td->files_size <= td->files_index) {
1581                 unsigned int new_size = td->o.nr_files + 1;
1582
1583                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1584
1585                 td->files = realloc(td->files, new_size * sizeof(f));
1586                 if (td->files == NULL) {
1587                         log_err("fio: realloc OOM\n");
1588                         assert(0);
1589                 }
1590                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1591                         td->file_locks = realloc(td->file_locks, new_size);
1592                         if (!td->file_locks) {
1593                                 log_err("fio: realloc OOM\n");
1594                                 assert(0);
1595                         }
1596                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1597                 }
1598                 td->files_size = new_size;
1599         }
1600         td->files[cur_files] = f;
1601         f->fileno = cur_files;
1602
1603         /*
1604          * init function, io engine may not be loaded yet
1605          */
1606         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1607                 f->real_file_size = -1ULL;
1608
1609         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1610                 f->file_name = strdup(file_name);
1611         else
1612                 f->file_name = smalloc_strdup(file_name);
1613         if (!f->file_name)
1614                 assert(0);
1615
1616         get_file_type(f);
1617
1618         switch (td->o.file_lock_mode) {
1619         case FILE_LOCK_NONE:
1620                 break;
1621         case FILE_LOCK_READWRITE:
1622                 f->rwlock = fio_rwlock_init();
1623                 break;
1624         case FILE_LOCK_EXCLUSIVE:
1625                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1626                 break;
1627         default:
1628                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1629                 assert(0);
1630         }
1631
1632         td->files_index++;
1633
1634         if (td->o.numjobs > 1)
1635                 set_already_allocated(file_name);
1636
1637         if (inc)
1638                 td->o.nr_files++;
1639
1640         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1641                                                         cur_files);
1642
1643         return cur_files;
1644 }
1645
1646 int add_file_exclusive(struct thread_data *td, const char *fname)
1647 {
1648         struct fio_file *f;
1649         unsigned int i;
1650
1651         for_each_file(td, f, i) {
1652                 if (!strcmp(f->file_name, fname))
1653                         return i;
1654         }
1655
1656         return add_file(td, fname, 0, 1);
1657 }
1658
1659 void get_file(struct fio_file *f)
1660 {
1661         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1662         assert(fio_file_open(f));
1663         f->references++;
1664 }
1665
1666 int put_file(struct thread_data *td, struct fio_file *f)
1667 {
1668         int f_ret = 0, ret = 0;
1669
1670         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1671
1672         if (!fio_file_open(f)) {
1673                 assert(f->fd == -1);
1674                 return 0;
1675         }
1676
1677         assert(f->references);
1678         if (--f->references)
1679                 return 0;
1680
1681         if (should_fsync(td) && td->o.fsync_on_close) {
1682                 f_ret = fsync(f->fd);
1683                 if (f_ret < 0)
1684                         f_ret = errno;
1685         }
1686
1687         if (td->io_ops->close_file)
1688                 ret = td->io_ops->close_file(td, f);
1689
1690         if (!ret)
1691                 ret = f_ret;
1692
1693         td->nr_open_files--;
1694         fio_file_clear_open(f);
1695         assert(f->fd == -1);
1696         return ret;
1697 }
1698
1699 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1700 {
1701         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1702                 return;
1703
1704         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1705                 if (ddir == DDIR_READ)
1706                         fio_rwlock_read(f->rwlock);
1707                 else
1708                         fio_rwlock_write(f->rwlock);
1709         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1710                 fio_sem_down(f->lock);
1711
1712         td->file_locks[f->fileno] = td->o.file_lock_mode;
1713 }
1714
1715 void unlock_file(struct thread_data *td, struct fio_file *f)
1716 {
1717         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1718                 return;
1719
1720         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1721                 fio_rwlock_unlock(f->rwlock);
1722         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1723                 fio_sem_up(f->lock);
1724
1725         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1726 }
1727
1728 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1729 {
1730         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1731                 return;
1732         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1733                 unlock_file(td, f);
1734 }
1735
1736 static bool recurse_dir(struct thread_data *td, const char *dirname)
1737 {
1738         struct dirent *dir;
1739         bool ret = false;
1740         DIR *D;
1741
1742         D = opendir(dirname);
1743         if (!D) {
1744                 char buf[FIO_VERROR_SIZE];
1745
1746                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1747                 td_verror(td, errno, buf);
1748                 return true;
1749         }
1750
1751         while ((dir = readdir(D)) != NULL) {
1752                 char full_path[PATH_MAX];
1753                 struct stat sb;
1754
1755                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1756                         continue;
1757
1758                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1759
1760                 if (lstat(full_path, &sb) == -1) {
1761                         if (errno != ENOENT) {
1762                                 td_verror(td, errno, "stat");
1763                                 ret = true;
1764                                 break;
1765                         }
1766                 }
1767
1768                 if (S_ISREG(sb.st_mode)) {
1769                         add_file(td, full_path, 0, 1);
1770                         continue;
1771                 }
1772                 if (!S_ISDIR(sb.st_mode))
1773                         continue;
1774
1775                 ret = recurse_dir(td, full_path);
1776                 if (ret)
1777                         break;
1778         }
1779
1780         closedir(D);
1781         return ret;
1782 }
1783
1784 int add_dir_files(struct thread_data *td, const char *path)
1785 {
1786         int ret = recurse_dir(td, path);
1787
1788         if (!ret)
1789                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1790
1791         return ret;
1792 }
1793
1794 void dup_files(struct thread_data *td, struct thread_data *org)
1795 {
1796         struct fio_file *f;
1797         unsigned int i;
1798
1799         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1800
1801         if (!org->files)
1802                 return;
1803
1804         td->files = malloc(org->files_index * sizeof(f));
1805
1806         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1807                 td->file_locks = malloc(org->files_index);
1808
1809         for_each_file(org, f, i) {
1810                 struct fio_file *__f;
1811
1812                 __f = alloc_new_file(td);
1813
1814                 if (f->file_name) {
1815                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1816                                 __f->file_name = strdup(f->file_name);
1817                         else
1818                                 __f->file_name = smalloc_strdup(f->file_name);
1819                         if (!__f->file_name)
1820                                 assert(0);
1821
1822                         __f->filetype = f->filetype;
1823                 }
1824
1825                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1826                         __f->lock = f->lock;
1827                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1828                         __f->rwlock = f->rwlock;
1829
1830                 td->files[i] = __f;
1831         }
1832 }
1833
1834 /*
1835  * Returns the index that matches the filename, or -1 if not there
1836  */
1837 int get_fileno(struct thread_data *td, const char *fname)
1838 {
1839         struct fio_file *f;
1840         unsigned int i;
1841
1842         for_each_file(td, f, i)
1843                 if (!strcmp(f->file_name, fname))
1844                         return i;
1845
1846         return -1;
1847 }
1848
1849 /*
1850  * For log usage, where we add/open/close files automatically
1851  */
1852 void free_release_files(struct thread_data *td)
1853 {
1854         close_files(td);
1855         td->o.nr_files = 0;
1856         td->o.open_files = 0;
1857         td->files_index = 0;
1858 }
1859
1860 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1861 {
1862         int i;
1863
1864         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1865                 f->last_pos[i] = f->file_offset;
1866                 f->last_start[i] = -1ULL;
1867         }
1868
1869         if (fio_file_axmap(f))
1870                 axmap_reset(f->io_axmap);
1871         else if (fio_file_lfsr(f))
1872                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1873 }
1874
1875 bool fio_files_done(struct thread_data *td)
1876 {
1877         struct fio_file *f;
1878         unsigned int i;
1879
1880         for_each_file(td, f, i)
1881                 if (!fio_file_done(f))
1882                         return false;
1883
1884         return true;
1885 }
1886
1887 /* free memory used in initialization phase only */
1888 void filesetup_mem_free(void)
1889 {
1890         free_already_allocated();
1891 }
1892
1893 /*
1894  * This function is for platforms which support direct I/O but not O_DIRECT.
1895  */
1896 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1897 {
1898 #ifdef FIO_OS_DIRECTIO
1899         int ret = fio_set_odirect(f);
1900
1901         if (ret) {
1902                 td_verror(td, ret, "fio_set_directio");
1903 #if defined(__sun__)
1904                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1905                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
1906                 } else {
1907                         log_err("fio: the file system does not seem to support direct IO\n");
1908                 }
1909 #else
1910                 log_err("fio: the file system does not seem to support direct IO\n");
1911 #endif
1912                 return -1;
1913         }
1914
1915         return 0;
1916 #else
1917         log_err("fio: direct IO is not supported on this host operating system\n");
1918         return -1;
1919 #endif
1920 }