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