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 "os/os.h"
  15
  16 #ifdef FIO_HAVE_LINUX_FALLOCATE
  17 #include <linux/falloc.h>
  18 #endif
  19
  20 static int root_warn;
  21
  22 static inline void clear_error(struct thread_data *td)
  23 {
  24         td->error = 0;
  25         td->verror[0] = '\0';
  26 }
  27
  28 /*
  29  * Leaves f->fd open on success, caller must close
  30  */
  31 static int extend_file(struct thread_data *td, struct fio_file *f)
  32 {
  33         int r, new_layout = 0, unlink_file = 0, flags;
  34         unsigned long long left;
  35         unsigned int bs;
  36         char *b;
  37
  38         if (read_only) {
  39                 log_err("fio: refusing extend of file due to read-only\n");
  40                 return 0;
  41         }
  42
  43         /*
  44          * check if we need to lay the file out complete again. fio
  45          * does that for operations involving reads, or for writes
  46          * where overwrite is set
  47          */
  48         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
  49             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
  50                 new_layout = 1;
  51         if (td_write(td) && !td->o.overwrite)
  52                 unlink_file = 1;
  53
  54         if (unlink_file || new_layout) {
  55                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  56                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
  57                         td_verror(td, errno, "unlink");
  58                         return 1;
  59                 }
  60         }
  61
  62         flags = O_WRONLY | O_CREAT;
  63         if (new_layout)
  64                 flags |= O_TRUNC;
  65
  66         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  67         f->fd = open(f->file_name, flags, 0644);
  68         if (f->fd < 0) {
  69                 td_verror(td, errno, "open");
  70                 return 1;
  71         }
  72
  73 #ifdef FIO_HAVE_FALLOCATE
  74         if (!td->o.fill_device) {
  75                 switch (td->o.fallocate_mode) {
  76                 case FIO_FALLOCATE_NONE:
  77                         break;
  78                 case FIO_FALLOCATE_POSIX:
  79                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  80                                  f->file_name, 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",
  85                                                 strerror(r));
  86                         }
  87                         break;
  88 #ifdef FIO_HAVE_LINUX_FALLOCATE
  89                 case FIO_FALLOCATE_KEEP_SIZE:
  90                         dprint(FD_FILE,
  91                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
  92                                 "file %s size %llu\n",
  93                                 f->file_name, f->real_file_size);
  94
  95                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
  96                                         f->real_file_size);
  97                         if (r != 0) {
  98                                 td_verror(td, errno, "fallocate");
  99                         }
 100                         break;
 101 #endif /* FIO_HAVE_LINUX_FALLOCATE */
 102                 default:
 103                         log_err("fio: unknown fallocate mode: %d\n",
 104                                 td->o.fallocate_mode);
 105                         assert(0);
 106                 }
 107         }
 108 #endif /* FIO_HAVE_FALLOCATE */
 109
 110         if (!new_layout)
 111                 goto done;
 112
 113         /*
 114          * The size will be -1ULL when fill_device is used, so don't truncate
 115          * or fallocate this file, just write it
 116          */
 117         if (!td->o.fill_device) {
 118                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 119                                                         f->real_file_size);
 120                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 121                         td_verror(td, errno, "ftruncate");
 122                         goto err;
 123                 }
 124         }
 125
 126         b = malloc(td->o.max_bs[DDIR_WRITE]);
 127         memset(b, 0, td->o.max_bs[DDIR_WRITE]);
 128
 129         left = f->real_file_size;
 130         while (left && !td->terminate) {
 131                 bs = td->o.max_bs[DDIR_WRITE];
 132                 if (bs > left)
 133                         bs = left;
 134
 135                 r = write(f->fd, b, bs);
 136
 137                 if (r > 0) {
 138                         left -= r;
 139                         continue;
 140                 } else {
 141                         if (r < 0) {
 142                                 int __e = errno;
 143
 144                                 if (__e == ENOSPC) {
 145                                         if (td->o.fill_device)
 146                                                 break;
 147                                         log_info("fio: ENOSPC on laying out "
 148                                                  "file, stopping\n");
 149                                         break;
 150                                 }
 151                                 td_verror(td, errno, "write");
 152                         } else
 153                                 td_verror(td, EIO, "write");
 154
 155                         break;
 156                 }
 157         }
 158
 159         if (td->terminate) {
 160                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 161                 unlink(f->file_name);
 162         } else if (td->o.create_fsync) {
 163                 if (fsync(f->fd) < 0) {
 164                         td_verror(td, errno, "fsync");
 165                         goto err;
 166                 }
 167         }
 168         if (td->o.fill_device && !td_write(td)) {
 169                 fio_file_clear_size_known(f);
 170                 if (td_io_get_file_size(td, f))
 171                         goto err;
 172                 if (f->io_size > f->real_file_size)
 173                         f->io_size = f->real_file_size;
 174         }
 175
 176         free(b);
 177 done:
 178         return 0;
 179 err:
 180         close(f->fd);
 181         f->fd = -1;
 182         return 1;
 183 }
 184
 185 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 186 {
 187         int r, did_open = 0, old_runstate;
 188         unsigned long long left;
 189         unsigned int bs;
 190         char *b;
 191
 192         if (td->io_ops->flags & FIO_PIPEIO)
 193                 return 0;
 194
 195         if (!fio_file_open(f)) {
 196                 if (td->io_ops->open_file(td, f)) {
 197                         log_err("fio: cannot pre-read, failed to open file\n");
 198                         return 1;
 199                 }
 200                 did_open = 1;
 201         }
 202
 203         old_runstate = td->runstate;
 204         td_set_runstate(td, TD_PRE_READING);
 205
 206         bs = td->o.max_bs[DDIR_READ];
 207         b = malloc(bs);
 208         memset(b, 0, bs);
 209
 210         lseek(f->fd, f->file_offset, SEEK_SET);
 211         left = f->io_size;
 212
 213         while (left && !td->terminate) {
 214                 if (bs > left)
 215                         bs = left;
 216
 217                 r = read(f->fd, b, bs);
 218
 219                 if (r == (int) bs) {
 220                         left -= bs;
 221                         continue;
 222                 } else {
 223                         td_verror(td, EIO, "pre_read");
 224                         break;
 225                 }
 226         }
 227
 228         td_set_runstate(td, old_runstate);
 229
 230         if (did_open)
 231                 td->io_ops->close_file(td, f);
 232         free(b);
 233         return 0;
 234 }
 235
 236 static unsigned long long get_rand_file_size(struct thread_data *td)
 237 {
 238         unsigned long long ret, sized;
 239         unsigned long r;
 240
 241         if (td->o.use_os_rand) {
 242                 r = os_random_long(&td->file_size_state);
 243                 sized = td->o.file_size_high - td->o.file_size_low;
 244                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 245         } else {
 246                 r = __rand(&td->__file_size_state);
 247                 sized = td->o.file_size_high - td->o.file_size_low;
 248                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 249         }
 250
 251         ret += td->o.file_size_low;
 252         ret -= (ret % td->o.rw_min_bs);
 253         return ret;
 254 }
 255
 256 static int file_size(struct thread_data *td, struct fio_file *f)
 257 {
 258         struct stat st;
 259
 260         if (stat(f->file_name, &st) == -1) {
 261                 td_verror(td, errno, "fstat");
 262                 return 1;
 263         }
 264
 265         f->real_file_size = st.st_size;
 266         return 0;
 267 }
 268
 269 static int bdev_size(struct thread_data *td, struct fio_file *f)
 270 {
 271         unsigned long long bytes = 0;
 272         int r;
 273
 274         if (td->io_ops->open_file(td, f)) {
 275                 log_err("fio: failed opening blockdev %s for size check\n",
 276                         f->file_name);
 277                 return 1;
 278         }
 279
 280         r = blockdev_size(f, &bytes);
 281         if (r) {
 282                 td_verror(td, r, "blockdev_size");
 283                 goto err;
 284         }
 285
 286         if (!bytes) {
 287                 log_err("%s: zero sized block device?\n", f->file_name);
 288                 goto err;
 289         }
 290
 291         f->real_file_size = bytes;
 292         td->io_ops->close_file(td, f);
 293         return 0;
 294 err:
 295         td->io_ops->close_file(td, f);
 296         return 1;
 297 }
 298
 299 static int char_size(struct thread_data *td, struct fio_file *f)
 300 {
 301 #ifdef FIO_HAVE_CHARDEV_SIZE
 302         unsigned long long bytes = 0;
 303         int r;
 304
 305         if (td->io_ops->open_file(td, f)) {
 306                 log_err("fio: failed opening blockdev %s for size check\n",
 307                         f->file_name);
 308                 return 1;
 309         }
 310
 311         r = chardev_size(f, &bytes);
 312         if (r) {
 313                 td_verror(td, r, "chardev_size");
 314                 goto err;
 315         }
 316
 317         if (!bytes) {
 318                 log_err("%s: zero sized char device?\n", f->file_name);
 319                 goto err;
 320         }
 321
 322         f->real_file_size = bytes;
 323         td->io_ops->close_file(td, f);
 324         return 0;
 325 err:
 326         td->io_ops->close_file(td, f);
 327         return 1;
 328 #else
 329         f->real_file_size = -1ULL;
 330         return 0;
 331 #endif
 332 }
 333
 334 static int get_file_size(struct thread_data *td, struct fio_file *f)
 335 {
 336         int ret = 0;
 337
 338         if (fio_file_size_known(f))
 339                 return 0;
 340
 341         if (f->filetype == FIO_TYPE_FILE)
 342                 ret = file_size(td, f);
 343         else if (f->filetype == FIO_TYPE_BD)
 344                 ret = bdev_size(td, f);
 345         else if (f->filetype == FIO_TYPE_CHAR)
 346                 ret = char_size(td, f);
 347         else
 348                 f->real_file_size = -1;
 349
 350         if (ret)
 351                 return ret;
 352
 353         if (f->file_offset > f->real_file_size) {
 354                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 355                                         f->file_offset, f->real_file_size);
 356                 return 1;
 357         }
 358
 359         fio_file_set_size_known(f);
 360         return 0;
 361 }
 362
 363 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 364                                    unsigned long long off,
 365                                    unsigned long long len)
 366 {
 367         int ret = 0;
 368
 369         if (len == -1ULL)
 370                 len = f->io_size;
 371         if (off == -1ULL)
 372                 off = f->file_offset;
 373
 374         if (len == -1ULL || off == -1ULL)
 375                 return 0;
 376
 377         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 378                                                                 len);
 379
 380         /*
 381          * FIXME: add blockdev flushing too
 382          */
 383         if (f->mmap_ptr) {
 384                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 385 #ifdef FIO_MADV_FREE
 386                 (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 387 #endif
 388         } else if (f->filetype == FIO_TYPE_FILE) {
 389                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 390         } else if (f->filetype == FIO_TYPE_BD) {
 391                 ret = blockdev_invalidate_cache(f);
 392                 if (ret < 0 && errno == EACCES && geteuid()) {
 393                         if (!root_warn) {
 394                                 log_err("fio: only root may flush block "
 395                                         "devices. Cache flush bypassed!\n");
 396                                 root_warn = 1;
 397                         }
 398                         ret = 0;
 399                 }
 400         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 401                 ret = 0;
 402
 403         if (ret < 0) {
 404                 td_verror(td, errno, "invalidate_cache");
 405                 return 1;
 406         } else if (ret > 0) {
 407                 td_verror(td, ret, "invalidate_cache");
 408                 return 1;
 409         }
 410
 411         return ret;
 412
 413 }
 414
 415 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 416 {
 417         if (!fio_file_open(f))
 418                 return 0;
 419
 420         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 421 }
 422
 423 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 424 {
 425         int ret = 0;
 426
 427         dprint(FD_FILE, "fd close %s\n", f->file_name);
 428
 429         remove_file_hash(f);
 430
 431         if (close(f->fd) < 0)
 432                 ret = errno;
 433
 434         f->fd = -1;
 435         return ret;
 436 }
 437
 438 int file_lookup_open(struct fio_file *f, int flags)
 439 {
 440         struct fio_file *__f;
 441         int from_hash;
 442
 443         __f = lookup_file_hash(f->file_name);
 444         if (__f) {
 445                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 446                 /*
 447                  * racy, need the __f->lock locked
 448                  */
 449                 f->lock = __f->lock;
 450                 f->lock_owner = __f->lock_owner;
 451                 f->lock_batch = __f->lock_batch;
 452                 f->lock_ddir = __f->lock_ddir;
 453                 from_hash = 1;
 454         } else {
 455                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 456                 from_hash = 0;
 457         }
 458
 459         f->fd = open(f->file_name, flags, 0600);
 460         return from_hash;
 461 }
 462
 463 int generic_open_file(struct thread_data *td, struct fio_file *f)
 464 {
 465         int is_std = 0;
 466         int flags = 0;
 467         int from_hash = 0;
 468
 469         dprint(FD_FILE, "fd open %s\n", f->file_name);
 470
 471         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 472                 log_err("fio: trim only applies to block device\n");
 473                 return 1;
 474         }
 475
 476         if (!strcmp(f->file_name, "-")) {
 477                 if (td_rw(td)) {
 478                         log_err("fio: can't read/write to stdin/out\n");
 479                         return 1;
 480                 }
 481                 is_std = 1;
 482
 483                 /*
 484                  * move output logging to stderr, if we are writing to stdout
 485                  */
 486                 if (td_write(td))
 487                         f_out = stderr;
 488         }
 489
 490         if (td_trim(td))
 491                 goto skip_flags;
 492         if (td->o.odirect)
 493                 flags |= OS_O_DIRECT;
 494         if (td->o.sync_io)
 495                 flags |= O_SYNC;
 496         if (td->o.create_on_open)
 497                 flags |= O_CREAT;
 498 skip_flags:
 499         if (f->filetype != FIO_TYPE_FILE)
 500                 flags |= FIO_O_NOATIME;
 501
 502 open_again:
 503         if (td_write(td)) {
 504                 if (!read_only)
 505                         flags |= O_RDWR;
 506
 507                 if (f->filetype == FIO_TYPE_FILE)
 508                         flags |= O_CREAT;
 509
 510                 if (is_std)
 511                         f->fd = dup(STDOUT_FILENO);
 512                 else
 513                         from_hash = file_lookup_open(f, flags);
 514         } else if (td_read(td)) {
 515                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 516                         flags |= O_RDWR;
 517                 else
 518                         flags |= O_RDONLY;
 519
 520                 if (is_std)
 521                         f->fd = dup(STDIN_FILENO);
 522                 else
 523                         from_hash = file_lookup_open(f, flags);
 524         } else { //td trim
 525                 flags |= O_RDWR;
 526                 from_hash = file_lookup_open(f, flags);
 527         }
 528
 529         if (f->fd == -1) {
 530                 char buf[FIO_VERROR_SIZE];
 531                 int __e = errno;
 532
 533                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 534                         flags &= ~FIO_O_NOATIME;
 535                         goto open_again;
 536                 }
 537
 538                 snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
 539
 540                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 541                         log_err("fio: looks like your file system does not " \
 542                                 "support direct=1/buffered=0\n");
 543                 }
 544
 545                 td_verror(td, __e, buf);
 546         }
 547
 548         if (!from_hash && f->fd != -1) {
 549                 if (add_file_hash(f)) {
 550                         int fio_unused ret;
 551
 552                         /*
 553                          * OK to ignore, we haven't done anything with it
 554                          */
 555                         ret = generic_close_file(td, f);
 556                         goto open_again;
 557                 }
 558         }
 559
 560         return 0;
 561 }
 562
 563 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 564 {
 565         return get_file_size(td, f);
 566 }
 567
 568 /*
 569  * open/close all files, so that ->real_file_size gets set
 570  */
 571 static int get_file_sizes(struct thread_data *td)
 572 {
 573         struct fio_file *f;
 574         unsigned int i;
 575         int err = 0;
 576
 577         for_each_file(td, f, i) {
 578                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 579                                                                 f->file_name);
 580
 581                 if (td_io_get_file_size(td, f)) {
 582                         if (td->error != ENOENT) {
 583                                 log_err("%s\n", td->verror);
 584                                 err = 1;
 585                         }
 586                         clear_error(td);
 587                 }
 588
 589                 if (f->real_file_size == -1ULL && td->o.size)
 590                         f->real_file_size = td->o.size / td->o.nr_files;
 591         }
 592
 593         return err;
 594 }
 595
 596 struct fio_mount {
 597         struct flist_head list;
 598         const char *base;
 599         char __base[256];
 600         unsigned int key;
 601 };
 602
 603 /*
 604  * Get free number of bytes for each file on each unique mount.
 605  */
 606 static unsigned long long get_fs_free_counts(struct thread_data *td)
 607 {
 608         struct flist_head *n, *tmp;
 609         unsigned long long ret = 0;
 610         struct fio_mount *fm;
 611         FLIST_HEAD(list);
 612         struct fio_file *f;
 613         unsigned int i;
 614
 615         for_each_file(td, f, i) {
 616                 struct stat sb;
 617                 char buf[256];
 618
 619                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 620                         if (f->real_file_size != -1ULL)
 621                                 ret += f->real_file_size;
 622                         continue;
 623                 } else if (f->filetype != FIO_TYPE_FILE)
 624                         continue;
 625
 626                 strcpy(buf, f->file_name);
 627
 628                 if (stat(buf, &sb) < 0) {
 629                         if (errno != ENOENT)
 630                                 break;
 631                         strcpy(buf, ".");
 632                         if (stat(buf, &sb) < 0)
 633                                 break;
 634                 }
 635
 636                 fm = NULL;
 637                 flist_for_each(n, &list) {
 638                         fm = flist_entry(n, struct fio_mount, list);
 639                         if (fm->key == sb.st_dev)
 640                                 break;
 641
 642                         fm = NULL;
 643                 }
 644
 645                 if (fm)
 646                         continue;
 647
 648                 fm = malloc(sizeof(*fm));
 649                 strcpy(fm->__base, buf);
 650                 fm->base = basename(fm->__base);
 651                 fm->key = sb.st_dev;
 652                 flist_add(&fm->list, &list);
 653         }
 654
 655         flist_for_each_safe(n, tmp, &list) {
 656                 unsigned long long sz;
 657
 658                 fm = flist_entry(n, struct fio_mount, list);
 659                 flist_del(&fm->list);
 660
 661                 sz = get_fs_size(fm->base);
 662                 if (sz && sz != -1ULL)
 663                         ret += sz;
 664
 665                 free(fm);
 666         }
 667
 668         return ret;
 669 }
 670
 671 unsigned long long get_start_offset(struct thread_data *td)
 672 {
 673         return td->o.start_offset +
 674                 (td->thread_number - 1) * td->o.offset_increment;
 675 }
 676
 677 /*
 678  * Open the files and setup files sizes, creating files if necessary.
 679  */
 680 int setup_files(struct thread_data *td)
 681 {
 682         unsigned long long total_size, extend_size;
 683         struct fio_file *f;
 684         unsigned int i;
 685         int err = 0, need_extend;
 686
 687         dprint(FD_FILE, "setup files\n");
 688
 689         if (td->o.read_iolog_file)
 690                 goto done;
 691
 692         /*
 693          * if ioengine defines a setup() method, it's responsible for
 694          * opening the files and setting f->real_file_size to indicate
 695          * the valid range for that file.
 696          */
 697         if (td->io_ops->setup)
 698                 err = td->io_ops->setup(td);
 699         else
 700                 err = get_file_sizes(td);
 701
 702         if (err)
 703                 return err;
 704
 705         /*
 706          * check sizes. if the files/devices do not exist and the size
 707          * isn't passed to fio, abort.
 708          */
 709         total_size = 0;
 710         for_each_file(td, f, i) {
 711                 if (f->real_file_size == -1ULL)
 712                         total_size = -1ULL;
 713                 else
 714                         total_size += f->real_file_size;
 715         }
 716
 717         if (td->o.fill_device)
 718                 td->fill_device_size = get_fs_free_counts(td);
 719
 720         /*
 721          * device/file sizes are zero and no size given, punt
 722          */
 723         if ((!total_size || total_size == -1ULL) && !td->o.size &&
 724             !(td->io_ops->flags & FIO_NOIO) && !td->o.fill_device) {
 725                 log_err("%s: you need to specify size=\n", td->o.name);
 726                 td_verror(td, EINVAL, "total_file_size");
 727                 return 1;
 728         }
 729
 730         /*
 731          * now file sizes are known, so we can set ->io_size. if size= is
 732          * not given, ->io_size is just equal to ->real_file_size. if size
 733          * is given, ->io_size is size / nr_files.
 734          */
 735         extend_size = total_size = 0;
 736         need_extend = 0;
 737         for_each_file(td, f, i) {
 738                 f->file_offset = get_start_offset(td);
 739
 740                 if (!td->o.file_size_low) {
 741                         /*
 742                          * no file size range given, file size is equal to
 743                          * total size divided by number of files. if that is
 744                          * zero, set it to the real file size.
 745                          */
 746                         f->io_size = td->o.size / td->o.nr_files;
 747                         if (!f->io_size)
 748                                 f->io_size = f->real_file_size - f->file_offset;
 749                 } else if (f->real_file_size < td->o.file_size_low ||
 750                            f->real_file_size > td->o.file_size_high) {
 751                         if (f->file_offset > td->o.file_size_low)
 752                                 goto err_offset;
 753                         /*
 754                          * file size given. if it's fixed, use that. if it's a
 755                          * range, generate a random size in-between.
 756                          */
 757                         if (td->o.file_size_low == td->o.file_size_high) {
 758                                 f->io_size = td->o.file_size_low
 759                                                 - f->file_offset;
 760                         } else {
 761                                 f->io_size = get_rand_file_size(td)
 762                                                 - f->file_offset;
 763                         }
 764                 } else
 765                         f->io_size = f->real_file_size - f->file_offset;
 766
 767                 if (f->io_size == -1ULL)
 768                         total_size = -1ULL;
 769                 else {
 770                         if (td->o.size_percent)
 771                                 f->io_size = (f->io_size * td->o.size_percent) / 100;
 772                         total_size += f->io_size;
 773                 }
 774
 775                 if (f->filetype == FIO_TYPE_FILE &&
 776                     (f->io_size + f->file_offset) > f->real_file_size &&
 777                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 778                         if (!td->o.create_on_open) {
 779                                 need_extend++;
 780                                 extend_size += (f->io_size + f->file_offset);
 781                         } else
 782                                 f->real_file_size = f->io_size + f->file_offset;
 783                         fio_file_set_extend(f);
 784                 }
 785         }
 786
 787         if (!td->o.size || td->o.size > total_size)
 788                 td->o.size = total_size;
 789
 790         /*
 791          * See if we need to extend some files
 792          */
 793         if (need_extend) {
 794                 temp_stall_ts = 1;
 795                 if (output_format == FIO_OUTPUT_NORMAL)
 796                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 797                                  " %lluMB)\n", td->o.name, need_extend,
 798                                         extend_size >> 20);
 799
 800                 for_each_file(td, f, i) {
 801                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 802
 803                         if (!fio_file_extend(f))
 804                                 continue;
 805
 806                         assert(f->filetype == FIO_TYPE_FILE);
 807                         fio_file_clear_extend(f);
 808                         if (!td->o.fill_device) {
 809                                 old_len = f->real_file_size;
 810                                 extend_len = f->io_size + f->file_offset -
 811                                                 old_len;
 812                         }
 813                         f->real_file_size = (f->io_size + f->file_offset);
 814                         err = extend_file(td, f);
 815                         if (err)
 816                                 break;
 817
 818                         err = __file_invalidate_cache(td, f, old_len,
 819                                                                 extend_len);
 820                         close(f->fd);
 821                         f->fd = -1;
 822                         if (err)
 823                                 break;
 824                 }
 825                 temp_stall_ts = 0;
 826         }
 827
 828         if (err)
 829                 return err;
 830
 831         if (!td->o.zone_size)
 832                 td->o.zone_size = td->o.size;
 833
 834         /*
 835          * iolog already set the total io size, if we read back
 836          * stored entries.
 837          */
 838         if (!td->o.read_iolog_file)
 839                 td->total_io_size = td->o.size * td->o.loops;
 840
 841 done:
 842         if (td->o.create_only)
 843                 td->done = 1;
 844
 845         return 0;
 846 err_offset:
 847         log_err("%s: you need to specify valid offset=\n", td->o.name);
 848         return 1;
 849 }
 850
 851 int pre_read_files(struct thread_data *td)
 852 {
 853         struct fio_file *f;
 854         unsigned int i;
 855
 856         dprint(FD_FILE, "pre_read files\n");
 857
 858         for_each_file(td, f, i) {
 859                 pre_read_file(td, f);
 860         }
 861
 862         return 1;
 863 }
 864
 865 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 866 {
 867         unsigned int range_size;
 868         unsigned long nranges;
 869
 870         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 871
 872         nranges = (f->real_file_size + range_size - 1) / range_size;
 873
 874         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 875                 zipf_init(&f->zipf, nranges, td->o.zipf_theta);
 876         else
 877                 pareto_init(&f->zipf, nranges, td->o.pareto_h);
 878
 879         return 1;
 880 }
 881
 882 static int init_rand_distribution(struct thread_data *td)
 883 {
 884         struct fio_file *f;
 885         unsigned int i;
 886         int state;
 887
 888         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 889                 return 0;
 890
 891         state = td->runstate;
 892         td_set_runstate(td, TD_SETTING_UP);
 893         for_each_file(td, f, i)
 894                 __init_rand_distribution(td, f);
 895         td_set_runstate(td, state);
 896
 897         return 1;
 898 }
 899
 900 int init_random_map(struct thread_data *td)
 901 {
 902         unsigned long long blocks, num_maps;
 903         struct fio_file *f;
 904         unsigned int i;
 905
 906         if (init_rand_distribution(td))
 907                 return 0;
 908         if (td->o.norandommap || !td_random(td))
 909                 return 0;
 910
 911         for_each_file(td, f, i) {
 912                 blocks = (f->real_file_size + td->o.rw_min_bs - 1) /
 913                                 (unsigned long long) td->o.rw_min_bs;
 914                 num_maps = (blocks + BLOCKS_PER_MAP - 1) /
 915                                 (unsigned long long) BLOCKS_PER_MAP;
 916                 if (num_maps == (unsigned long) num_maps) {
 917                         f->file_map = smalloc(num_maps * sizeof(unsigned long));
 918                         if (f->file_map) {
 919                                 f->num_maps = num_maps;
 920                                 continue;
 921                         }
 922                 } else
 923                         f->file_map = NULL;
 924
 925                 if (!td->o.softrandommap) {
 926                         log_err("fio: failed allocating random map. If running"
 927                                 " a large number of jobs, try the 'norandommap'"
 928                                 " option or set 'softrandommap'. Or give"
 929                                 " a larger --alloc-size to fio.\n");
 930                         return 1;
 931                 }
 932
 933                 log_info("fio: file %s failed allocating random map. Running "
 934                          "job without.\n", f->file_name);
 935                 f->num_maps = 0;
 936         }
 937
 938         return 0;
 939 }
 940
 941 void close_files(struct thread_data *td)
 942 {
 943         struct fio_file *f;
 944         unsigned int i;
 945
 946         for_each_file(td, f, i) {
 947                 if (fio_file_open(f))
 948                         td_io_close_file(td, f);
 949         }
 950 }
 951
 952 void close_and_free_files(struct thread_data *td)
 953 {
 954         struct fio_file *f;
 955         unsigned int i;
 956
 957         dprint(FD_FILE, "close files\n");
 958
 959         for_each_file(td, f, i) {
 960                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
 961                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
 962                         unlink(f->file_name);
 963                 }
 964
 965                 if (fio_file_open(f))
 966                         td_io_close_file(td, f);
 967
 968                 remove_file_hash(f);
 969
 970                 sfree(f->file_name);
 971                 f->file_name = NULL;
 972                 sfree(f->file_map);
 973                 f->file_map = NULL;
 974                 sfree(f);
 975         }
 976
 977         td->o.filename = NULL;
 978         free(td->files);
 979         td->files_index = 0;
 980         td->files = NULL;
 981         td->o.nr_files = 0;
 982 }
 983
 984 static void get_file_type(struct fio_file *f)
 985 {
 986         struct stat sb;
 987
 988         if (!strcmp(f->file_name, "-"))
 989                 f->filetype = FIO_TYPE_PIPE;
 990         else
 991                 f->filetype = FIO_TYPE_FILE;
 992
 993         /* \\.\ is the device namespace in Windows, where every file is
 994          * a block device */
 995         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
 996                 f->filetype = FIO_TYPE_BD;
 997
 998         if (!stat(f->file_name, &sb)) {
 999                 if (S_ISBLK(sb.st_mode))
1000                         f->filetype = FIO_TYPE_BD;
1001                 else if (S_ISCHR(sb.st_mode))
1002                         f->filetype = FIO_TYPE_CHAR;
1003                 else if (S_ISFIFO(sb.st_mode))
1004                         f->filetype = FIO_TYPE_PIPE;
1005         }
1006 }
1007
1008 int add_file(struct thread_data *td, const char *fname)
1009 {
1010         int cur_files = td->files_index;
1011         char file_name[PATH_MAX];
1012         struct fio_file *f;
1013         int len = 0;
1014
1015         dprint(FD_FILE, "add file %s\n", fname);
1016
1017         f = smalloc(sizeof(*f));
1018         if (!f) {
1019                 log_err("fio: smalloc OOM\n");
1020                 assert(0);
1021         }
1022
1023         f->fd = -1;
1024         fio_file_reset(f);
1025
1026         if (td->files_size <= td->files_index) {
1027                 int new_size = td->o.nr_files + 1;
1028
1029                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1030
1031                 td->files = realloc(td->files, new_size * sizeof(f));
1032                 td->files_size = new_size;
1033         }
1034         td->files[cur_files] = f;
1035         f->fileno = cur_files;
1036
1037         /*
1038          * init function, io engine may not be loaded yet
1039          */
1040         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1041                 f->real_file_size = -1ULL;
1042
1043         if (td->o.directory)
1044                 len = sprintf(file_name, "%s/", td->o.directory);
1045
1046         sprintf(file_name + len, "%s", fname);
1047         f->file_name = smalloc_strdup(file_name);
1048         if (!f->file_name) {
1049                 log_err("fio: smalloc OOM\n");
1050                 assert(0);
1051         }
1052
1053         get_file_type(f);
1054
1055         switch (td->o.file_lock_mode) {
1056         case FILE_LOCK_NONE:
1057                 break;
1058         case FILE_LOCK_READWRITE:
1059                 f->lock = fio_mutex_rw_init();
1060                 break;
1061         case FILE_LOCK_EXCLUSIVE:
1062                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1063                 break;
1064         default:
1065                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1066                 assert(0);
1067         }
1068
1069         td->files_index++;
1070         if (f->filetype == FIO_TYPE_FILE)
1071                 td->nr_normal_files++;
1072
1073         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1074                                                         cur_files);
1075
1076         return cur_files;
1077 }
1078
1079 int add_file_exclusive(struct thread_data *td, const char *fname)
1080 {
1081         struct fio_file *f;
1082         unsigned int i;
1083
1084         for_each_file(td, f, i) {
1085                 if (!strcmp(f->file_name, fname))
1086                         return i;
1087         }
1088
1089         return add_file(td, fname);
1090 }
1091
1092 void get_file(struct fio_file *f)
1093 {
1094         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1095         assert(fio_file_open(f));
1096         f->references++;
1097 }
1098
1099 int put_file(struct thread_data *td, struct fio_file *f)
1100 {
1101         int f_ret = 0, ret = 0;
1102
1103         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1104
1105         if (!fio_file_open(f)) {
1106                 assert(f->fd == -1);
1107                 return 0;
1108         }
1109
1110         assert(f->references);
1111         if (--f->references)
1112                 return 0;
1113
1114         if (should_fsync(td) && td->o.fsync_on_close)
1115                 f_ret = fsync(f->fd);
1116
1117         if (td->io_ops->close_file)
1118                 ret = td->io_ops->close_file(td, f);
1119
1120         if (!ret)
1121                 ret = f_ret;
1122
1123         td->nr_open_files--;
1124         fio_file_clear_open(f);
1125         assert(f->fd == -1);
1126         return ret;
1127 }
1128
1129 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1130 {
1131         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1132                 return;
1133
1134         if (f->lock_owner == td && f->lock_batch--)
1135                 return;
1136
1137         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1138                 if (ddir == DDIR_READ)
1139                         fio_mutex_down_read(f->lock);
1140                 else
1141                         fio_mutex_down_write(f->lock);
1142         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1143                 fio_mutex_down(f->lock);
1144
1145         f->lock_owner = td;
1146         f->lock_batch = td->o.lockfile_batch;
1147         f->lock_ddir = ddir;
1148 }
1149
1150 void unlock_file(struct thread_data *td, struct fio_file *f)
1151 {
1152         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1153                 return;
1154         if (f->lock_batch)
1155                 return;
1156
1157         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1158                 const int is_read = f->lock_ddir == DDIR_READ;
1159                 int val = fio_mutex_getval(f->lock);
1160
1161                 if ((is_read && val == 1) || (!is_read && val == -1))
1162                         f->lock_owner = NULL;
1163
1164                 if (is_read)
1165                         fio_mutex_up_read(f->lock);
1166                 else
1167                         fio_mutex_up_write(f->lock);
1168         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) {
1169                 int val = fio_mutex_getval(f->lock);
1170
1171                 if (val == 0)
1172                         f->lock_owner = NULL;
1173
1174                 fio_mutex_up(f->lock);
1175         }
1176 }
1177
1178 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1179 {
1180         if (f->lock_owner != td)
1181                 return;
1182
1183         f->lock_batch = 0;
1184         unlock_file(td, f);
1185 }
1186
1187 static int recurse_dir(struct thread_data *td, const char *dirname)
1188 {
1189         struct dirent *dir;
1190         int ret = 0;
1191         DIR *D;
1192
1193         D = opendir(dirname);
1194         if (!D) {
1195                 char buf[FIO_VERROR_SIZE];
1196
1197                 snprintf(buf, FIO_VERROR_SIZE - 1, "opendir(%s)", dirname);
1198                 td_verror(td, errno, buf);
1199                 return 1;
1200         }
1201
1202         while ((dir = readdir(D)) != NULL) {
1203                 char full_path[PATH_MAX];
1204                 struct stat sb;
1205
1206                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1207                         continue;
1208
1209                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1210
1211                 if (lstat(full_path, &sb) == -1) {
1212                         if (errno != ENOENT) {
1213                                 td_verror(td, errno, "stat");
1214                                 return 1;
1215                         }
1216                 }
1217
1218                 if (S_ISREG(sb.st_mode)) {
1219                         add_file(td, full_path);
1220                         td->o.nr_files++;
1221                         continue;
1222                 }
1223                 if (!S_ISDIR(sb.st_mode))
1224                         continue;
1225
1226                 ret = recurse_dir(td, full_path);
1227                 if (ret)
1228                         break;
1229         }
1230
1231         closedir(D);
1232         return ret;
1233 }
1234
1235 int add_dir_files(struct thread_data *td, const char *path)
1236 {
1237         int ret = recurse_dir(td, path);
1238
1239         if (!ret)
1240                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1241
1242         return ret;
1243 }
1244
1245 void dup_files(struct thread_data *td, struct thread_data *org)
1246 {
1247         struct fio_file *f;
1248         unsigned int i;
1249
1250         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1251
1252         if (!org->files)
1253                 return;
1254
1255         td->files = malloc(org->files_index * sizeof(f));
1256
1257         for_each_file(org, f, i) {
1258                 struct fio_file *__f;
1259
1260                 __f = smalloc(sizeof(*__f));
1261                 if (!__f) {
1262                         log_err("fio: smalloc OOM\n");
1263                         assert(0);
1264                 }
1265                 __f->fd = -1;
1266                 fio_file_reset(__f);
1267
1268                 if (f->file_name) {
1269                         __f->file_name = smalloc_strdup(f->file_name);
1270                         if (!__f->file_name) {
1271                                 log_err("fio: smalloc OOM\n");
1272                                 assert(0);
1273                         }
1274
1275                         __f->filetype = f->filetype;
1276                 }
1277
1278                 td->files[i] = __f;
1279         }
1280 }
1281
1282 /*
1283  * Returns the index that matches the filename, or -1 if not there
1284  */
1285 int get_fileno(struct thread_data *td, const char *fname)
1286 {
1287         struct fio_file *f;
1288         unsigned int i;
1289
1290         for_each_file(td, f, i)
1291                 if (!strcmp(f->file_name, fname))
1292                         return i;
1293
1294         return -1;
1295 }
1296
1297 /*
1298  * For log usage, where we add/open/close files automatically
1299  */
1300 void free_release_files(struct thread_data *td)
1301 {
1302         close_files(td);
1303         td->files_index = 0;
1304         td->nr_normal_files = 0;
1305 }