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