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 #include "hash.h"
  16 #include "lib/axmap.h"
  17
  18 #ifdef FIO_HAVE_LINUX_FALLOCATE
  19 #include <linux/falloc.h>
  20 #endif
  21
  22 static int root_warn;
  23
  24 static inline void clear_error(struct thread_data *td)
  25 {
  26         td->error = 0;
  27         td->verror[0] = '\0';
  28 }
  29
  30 /*
  31  * Leaves f->fd open on success, caller must close
  32  */
  33 static int extend_file(struct thread_data *td, struct fio_file *f)
  34 {
  35         int r, new_layout = 0, unlink_file = 0, flags;
  36         unsigned long long left;
  37         unsigned int bs;
  38         char *b;
  39
  40         if (read_only) {
  41                 log_err("fio: refusing extend of file due to read-only\n");
  42                 return 0;
  43         }
  44
  45         /*
  46          * check if we need to lay the file out complete again. fio
  47          * does that for operations involving reads, or for writes
  48          * where overwrite is set
  49          */
  50         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
  51             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
  52                 new_layout = 1;
  53         if (td_write(td) && !td->o.overwrite)
  54                 unlink_file = 1;
  55
  56         if (unlink_file || new_layout) {
  57                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  58                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
  59                         td_verror(td, errno, "unlink");
  60                         return 1;
  61                 }
  62         }
  63
  64         flags = O_WRONLY | O_CREAT;
  65         if (new_layout)
  66                 flags |= O_TRUNC;
  67
  68         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  69         f->fd = open(f->file_name, flags, 0644);
  70         if (f->fd < 0) {
  71                 td_verror(td, errno, "open");
  72                 return 1;
  73         }
  74
  75 #ifdef FIO_HAVE_FALLOCATE
  76         if (!td->o.fill_device) {
  77                 switch (td->o.fallocate_mode) {
  78                 case FIO_FALLOCATE_NONE:
  79                         break;
  80                 case FIO_FALLOCATE_POSIX:
  81                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  82                                  f->file_name, f->real_file_size);
  83
  84                         r = posix_fallocate(f->fd, 0, f->real_file_size);
  85                         if (r > 0) {
  86                                 log_err("fio: posix_fallocate fails: %s\n",
  87                                                 strerror(r));
  88                         }
  89                         break;
  90 #ifdef FIO_HAVE_LINUX_FALLOCATE
  91                 case FIO_FALLOCATE_KEEP_SIZE:
  92                         dprint(FD_FILE,
  93                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
  94                                 "file %s size %llu\n",
  95                                 f->file_name, f->real_file_size);
  96
  97                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
  98                                         f->real_file_size);
  99                         if (r != 0)
 100                                 td_verror(td, errno, "fallocate");
 101
 102                         break;
 103 #endif /* FIO_HAVE_LINUX_FALLOCATE */
 104                 default:
 105                         log_err("fio: unknown fallocate mode: %d\n",
 106                                 td->o.fallocate_mode);
 107                         assert(0);
 108                 }
 109         }
 110 #endif /* FIO_HAVE_FALLOCATE */
 111
 112         if (!new_layout)
 113                 goto done;
 114
 115         /*
 116          * The size will be -1ULL when fill_device is used, so don't truncate
 117          * or fallocate this file, just write it
 118          */
 119         if (!td->o.fill_device) {
 120                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 121                                                         f->real_file_size);
 122                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 123                         td_verror(td, errno, "ftruncate");
 124                         goto err;
 125                 }
 126         }
 127
 128         b = malloc(td->o.max_bs[DDIR_WRITE]);
 129         memset(b, 0, td->o.max_bs[DDIR_WRITE]);
 130
 131         left = f->real_file_size;
 132         while (left && !td->terminate) {
 133                 bs = td->o.max_bs[DDIR_WRITE];
 134                 if (bs > left)
 135                         bs = left;
 136
 137                 r = write(f->fd, b, bs);
 138
 139                 if (r > 0) {
 140                         left -= r;
 141                         continue;
 142                 } else {
 143                         if (r < 0) {
 144                                 int __e = errno;
 145
 146                                 if (__e == ENOSPC) {
 147                                         if (td->o.fill_device)
 148                                                 break;
 149                                         log_info("fio: ENOSPC on laying out "
 150                                                  "file, stopping\n");
 151                                         break;
 152                                 }
 153                                 td_verror(td, errno, "write");
 154                         } else
 155                                 td_verror(td, EIO, "write");
 156
 157                         break;
 158                 }
 159         }
 160
 161         if (td->terminate) {
 162                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 163                 unlink(f->file_name);
 164         } else if (td->o.create_fsync) {
 165                 if (fsync(f->fd) < 0) {
 166                         td_verror(td, errno, "fsync");
 167                         goto err;
 168                 }
 169         }
 170         if (td->o.fill_device && !td_write(td)) {
 171                 fio_file_clear_size_known(f);
 172                 if (td_io_get_file_size(td, f))
 173                         goto err;
 174                 if (f->io_size > f->real_file_size)
 175                         f->io_size = f->real_file_size;
 176         }
 177
 178         free(b);
 179 done:
 180         return 0;
 181 err:
 182         close(f->fd);
 183         f->fd = -1;
 184         return 1;
 185 }
 186
 187 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 188 {
 189         int r, did_open = 0, old_runstate;
 190         unsigned long long left;
 191         unsigned int bs;
 192         char *b;
 193
 194         if (td->io_ops->flags & FIO_PIPEIO)
 195                 return 0;
 196
 197         if (!fio_file_open(f)) {
 198                 if (td->io_ops->open_file(td, f)) {
 199                         log_err("fio: cannot pre-read, failed to open file\n");
 200                         return 1;
 201                 }
 202                 did_open = 1;
 203         }
 204
 205         old_runstate = td->runstate;
 206         td_set_runstate(td, TD_PRE_READING);
 207
 208         bs = td->o.max_bs[DDIR_READ];
 209         b = malloc(bs);
 210         memset(b, 0, bs);
 211
 212         lseek(f->fd, f->file_offset, SEEK_SET);
 213         left = f->io_size;
 214
 215         while (left && !td->terminate) {
 216                 if (bs > left)
 217                         bs = left;
 218
 219                 r = read(f->fd, b, bs);
 220
 221                 if (r == (int) bs) {
 222                         left -= bs;
 223                         continue;
 224                 } else {
 225                         td_verror(td, EIO, "pre_read");
 226                         break;
 227                 }
 228         }
 229
 230         td_set_runstate(td, old_runstate);
 231
 232         if (did_open)
 233                 td->io_ops->close_file(td, f);
 234         free(b);
 235         return 0;
 236 }
 237
 238 static unsigned long long get_rand_file_size(struct thread_data *td)
 239 {
 240         unsigned long long ret, sized;
 241         unsigned long r;
 242
 243         if (td->o.use_os_rand) {
 244                 r = os_random_long(&td->file_size_state);
 245                 sized = td->o.file_size_high - td->o.file_size_low;
 246                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 247         } else {
 248                 r = __rand(&td->__file_size_state);
 249                 sized = td->o.file_size_high - td->o.file_size_low;
 250                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 251         }
 252
 253         ret += td->o.file_size_low;
 254         ret -= (ret % td->o.rw_min_bs);
 255         return ret;
 256 }
 257
 258 static int file_size(struct thread_data *td, struct fio_file *f)
 259 {
 260         struct stat st;
 261
 262         if (stat(f->file_name, &st) == -1) {
 263                 td_verror(td, errno, "fstat");
 264                 return 1;
 265         }
 266
 267         f->real_file_size = st.st_size;
 268         return 0;
 269 }
 270
 271 static int bdev_size(struct thread_data *td, struct fio_file *f)
 272 {
 273         unsigned long long bytes = 0;
 274         int r;
 275
 276         if (td->io_ops->open_file(td, f)) {
 277                 log_err("fio: failed opening blockdev %s for size check\n",
 278                         f->file_name);
 279                 return 1;
 280         }
 281
 282         r = blockdev_size(f, &bytes);
 283         if (r) {
 284                 td_verror(td, r, "blockdev_size");
 285                 goto err;
 286         }
 287
 288         if (!bytes) {
 289                 log_err("%s: zero sized block device?\n", f->file_name);
 290                 goto err;
 291         }
 292
 293         f->real_file_size = bytes;
 294         td->io_ops->close_file(td, f);
 295         return 0;
 296 err:
 297         td->io_ops->close_file(td, f);
 298         return 1;
 299 }
 300
 301 static int char_size(struct thread_data *td, struct fio_file *f)
 302 {
 303 #ifdef FIO_HAVE_CHARDEV_SIZE
 304         unsigned long long bytes = 0;
 305         int r;
 306
 307         if (td->io_ops->open_file(td, f)) {
 308                 log_err("fio: failed opening blockdev %s for size check\n",
 309                         f->file_name);
 310                 return 1;
 311         }
 312
 313         r = chardev_size(f, &bytes);
 314         if (r) {
 315                 td_verror(td, r, "chardev_size");
 316                 goto err;
 317         }
 318
 319         if (!bytes) {
 320                 log_err("%s: zero sized char device?\n", f->file_name);
 321                 goto err;
 322         }
 323
 324         f->real_file_size = bytes;
 325         td->io_ops->close_file(td, f);
 326         return 0;
 327 err:
 328         td->io_ops->close_file(td, f);
 329         return 1;
 330 #else
 331         f->real_file_size = -1ULL;
 332         return 0;
 333 #endif
 334 }
 335
 336 static int get_file_size(struct thread_data *td, struct fio_file *f)
 337 {
 338         int ret = 0;
 339
 340         if (fio_file_size_known(f))
 341                 return 0;
 342
 343         if (f->filetype == FIO_TYPE_FILE)
 344                 ret = file_size(td, f);
 345         else if (f->filetype == FIO_TYPE_BD)
 346                 ret = bdev_size(td, f);
 347         else if (f->filetype == FIO_TYPE_CHAR)
 348                 ret = char_size(td, f);
 349         else
 350                 f->real_file_size = -1;
 351
 352         if (ret)
 353                 return ret;
 354
 355         if (f->file_offset > f->real_file_size) {
 356                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 357                                         f->file_offset, f->real_file_size);
 358                 return 1;
 359         }
 360
 361         fio_file_set_size_known(f);
 362         return 0;
 363 }
 364
 365 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 366                                    unsigned long long off,
 367                                    unsigned long long len)
 368 {
 369         int ret = 0;
 370
 371         if (len == -1ULL)
 372                 len = f->io_size;
 373         if (off == -1ULL)
 374                 off = f->file_offset;
 375
 376         if (len == -1ULL || off == -1ULL)
 377                 return 0;
 378
 379         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 380                                                                 len);
 381
 382         /*
 383          * FIXME: add blockdev flushing too
 384          */
 385         if (f->mmap_ptr) {
 386                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 387 #ifdef FIO_MADV_FREE
 388                 (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 389 #endif
 390         } else if (f->filetype == FIO_TYPE_FILE) {
 391                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 392         } else if (f->filetype == FIO_TYPE_BD) {
 393                 ret = blockdev_invalidate_cache(f);
 394                 if (ret < 0 && errno == EACCES && geteuid()) {
 395                         if (!root_warn) {
 396                                 log_err("fio: only root may flush block "
 397                                         "devices. Cache flush bypassed!\n");
 398                                 root_warn = 1;
 399                         }
 400                         ret = 0;
 401                 }
 402         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 403                 ret = 0;
 404
 405         if (ret < 0) {
 406                 td_verror(td, errno, "invalidate_cache");
 407                 return 1;
 408         } else if (ret > 0) {
 409                 td_verror(td, ret, "invalidate_cache");
 410                 return 1;
 411         }
 412
 413         return ret;
 414
 415 }
 416
 417 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 418 {
 419         if (!fio_file_open(f))
 420                 return 0;
 421
 422         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 423 }
 424
 425 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 426 {
 427         int ret = 0;
 428
 429         dprint(FD_FILE, "fd close %s\n", f->file_name);
 430
 431         remove_file_hash(f);
 432
 433         if (close(f->fd) < 0)
 434                 ret = errno;
 435
 436         f->fd = -1;
 437
 438         if (f->shadow_fd != -1) {
 439                 close(f->shadow_fd);
 440                 f->shadow_fd = -1;
 441         }
 442
 443         return ret;
 444 }
 445
 446 int file_lookup_open(struct fio_file *f, int flags)
 447 {
 448         struct fio_file *__f;
 449         int from_hash;
 450
 451         __f = lookup_file_hash(f->file_name);
 452         if (__f) {
 453                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 454                 /*
 455                  * racy, need the __f->lock locked
 456                  */
 457                 f->lock = __f->lock;
 458                 f->lock_owner = __f->lock_owner;
 459                 f->lock_batch = __f->lock_batch;
 460                 f->lock_ddir = __f->lock_ddir;
 461                 from_hash = 1;
 462         } else {
 463                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 464                 from_hash = 0;
 465         }
 466
 467         f->fd = open(f->file_name, flags, 0600);
 468         return from_hash;
 469 }
 470
 471 static int file_close_shadow_fds(struct thread_data *td)
 472 {
 473         struct fio_file *f;
 474         int num_closed = 0;
 475         unsigned int i;
 476
 477         for_each_file(td, f, i) {
 478                 if (f->shadow_fd == -1)
 479                         continue;
 480
 481                 close(f->shadow_fd);
 482                 f->shadow_fd = -1;
 483                 num_closed++;
 484         }
 485
 486         return num_closed;
 487 }
 488
 489 int generic_open_file(struct thread_data *td, struct fio_file *f)
 490 {
 491         int is_std = 0;
 492         int flags = 0;
 493         int from_hash = 0;
 494
 495         dprint(FD_FILE, "fd open %s\n", f->file_name);
 496
 497         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 498                 log_err("fio: trim only applies to block device\n");
 499                 return 1;
 500         }
 501
 502         if (!strcmp(f->file_name, "-")) {
 503                 if (td_rw(td)) {
 504                         log_err("fio: can't read/write to stdin/out\n");
 505                         return 1;
 506                 }
 507                 is_std = 1;
 508
 509                 /*
 510                  * move output logging to stderr, if we are writing to stdout
 511                  */
 512                 if (td_write(td))
 513                         f_out = stderr;
 514         }
 515
 516         if (td_trim(td))
 517                 goto skip_flags;
 518         if (td->o.odirect)
 519                 flags |= OS_O_DIRECT;
 520         if (td->o.sync_io)
 521                 flags |= O_SYNC;
 522         if (td->o.create_on_open)
 523                 flags |= O_CREAT;
 524 skip_flags:
 525         if (f->filetype != FIO_TYPE_FILE)
 526                 flags |= FIO_O_NOATIME;
 527
 528 open_again:
 529         if (td_write(td)) {
 530                 if (!read_only)
 531                         flags |= O_RDWR;
 532
 533                 if (f->filetype == FIO_TYPE_FILE)
 534                         flags |= O_CREAT;
 535
 536                 if (is_std)
 537                         f->fd = dup(STDOUT_FILENO);
 538                 else
 539                         from_hash = file_lookup_open(f, flags);
 540         } else if (td_read(td)) {
 541                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 542                         flags |= O_RDWR;
 543                 else
 544                         flags |= O_RDONLY;
 545
 546                 if (is_std)
 547                         f->fd = dup(STDIN_FILENO);
 548                 else
 549                         from_hash = file_lookup_open(f, flags);
 550         } else { //td trim
 551                 flags |= O_RDWR;
 552                 from_hash = file_lookup_open(f, flags);
 553         }
 554
 555         if (f->fd == -1) {
 556                 char buf[FIO_VERROR_SIZE];
 557                 int __e = errno;
 558
 559                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 560                         flags &= ~FIO_O_NOATIME;
 561                         goto open_again;
 562                 }
 563                 if (__e == EMFILE && file_close_shadow_fds(td))
 564                         goto open_again;
 565
 566                 snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
 567
 568                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 569                         log_err("fio: looks like your file system does not " \
 570                                 "support direct=1/buffered=0\n");
 571                 }
 572
 573                 td_verror(td, __e, buf);
 574         }
 575
 576         if (!from_hash && f->fd != -1) {
 577                 if (add_file_hash(f)) {
 578                         int fio_unused ret;
 579
 580                         /*
 581                          * Stash away descriptor for later close. This is to
 582                          * work-around a "feature" on Linux, where a close of
 583                          * an fd that has been opened for write will trigger
 584                          * udev to call blkid to check partitions, fs id, etc.
 585                          * That polutes the device cache, which can slow down
 586                          * unbuffered accesses.
 587                          */
 588                         if (f->shadow_fd == -1)
 589                                 f->shadow_fd = f->fd;
 590                         else {
 591                                 /*
 592                                  * OK to ignore, we haven't done anything
 593                                  * with it
 594                                  */
 595                                 ret = generic_close_file(td, f);
 596                         }
 597                         goto open_again;
 598                 }
 599         }
 600
 601         return 0;
 602 }
 603
 604 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 605 {
 606         return get_file_size(td, f);
 607 }
 608
 609 /*
 610  * open/close all files, so that ->real_file_size gets set
 611  */
 612 static int get_file_sizes(struct thread_data *td)
 613 {
 614         struct fio_file *f;
 615         unsigned int i;
 616         int err = 0;
 617
 618         for_each_file(td, f, i) {
 619                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 620                                                                 f->file_name);
 621
 622                 if (td_io_get_file_size(td, f)) {
 623                         if (td->error != ENOENT) {
 624                                 log_err("%s\n", td->verror);
 625                                 err = 1;
 626                         }
 627                         clear_error(td);
 628                 }
 629
 630                 if (f->real_file_size == -1ULL && td->o.size)
 631                         f->real_file_size = td->o.size / td->o.nr_files;
 632         }
 633
 634         return err;
 635 }
 636
 637 struct fio_mount {
 638         struct flist_head list;
 639         const char *base;
 640         char __base[256];
 641         unsigned int key;
 642 };
 643
 644 /*
 645  * Get free number of bytes for each file on each unique mount.
 646  */
 647 static unsigned long long get_fs_free_counts(struct thread_data *td)
 648 {
 649         struct flist_head *n, *tmp;
 650         unsigned long long ret = 0;
 651         struct fio_mount *fm;
 652         FLIST_HEAD(list);
 653         struct fio_file *f;
 654         unsigned int i;
 655
 656         for_each_file(td, f, i) {
 657                 struct stat sb;
 658                 char buf[256];
 659
 660                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 661                         if (f->real_file_size != -1ULL)
 662                                 ret += f->real_file_size;
 663                         continue;
 664                 } else if (f->filetype != FIO_TYPE_FILE)
 665                         continue;
 666
 667                 strcpy(buf, f->file_name);
 668
 669                 if (stat(buf, &sb) < 0) {
 670                         if (errno != ENOENT)
 671                                 break;
 672                         strcpy(buf, ".");
 673                         if (stat(buf, &sb) < 0)
 674                                 break;
 675                 }
 676
 677                 fm = NULL;
 678                 flist_for_each(n, &list) {
 679                         fm = flist_entry(n, struct fio_mount, list);
 680                         if (fm->key == sb.st_dev)
 681                                 break;
 682
 683                         fm = NULL;
 684                 }
 685
 686                 if (fm)
 687                         continue;
 688
 689                 fm = malloc(sizeof(*fm));
 690                 strcpy(fm->__base, buf);
 691                 fm->base = basename(fm->__base);
 692                 fm->key = sb.st_dev;
 693                 flist_add(&fm->list, &list);
 694         }
 695
 696         flist_for_each_safe(n, tmp, &list) {
 697                 unsigned long long sz;
 698
 699                 fm = flist_entry(n, struct fio_mount, list);
 700                 flist_del(&fm->list);
 701
 702                 sz = get_fs_size(fm->base);
 703                 if (sz && sz != -1ULL)
 704                         ret += sz;
 705
 706                 free(fm);
 707         }
 708
 709         return ret;
 710 }
 711
 712 uint64_t get_start_offset(struct thread_data *td)
 713 {
 714         return td->o.start_offset +
 715                 (td->thread_number - 1) * td->o.offset_increment;
 716 }
 717
 718 /*
 719  * Open the files and setup files sizes, creating files if necessary.
 720  */
 721 int setup_files(struct thread_data *td)
 722 {
 723         unsigned long long total_size, extend_size;
 724         struct fio_file *f;
 725         unsigned int i;
 726         int err = 0, need_extend;
 727
 728         dprint(FD_FILE, "setup files\n");
 729
 730         if (td->o.read_iolog_file)
 731                 goto done;
 732
 733         /*
 734          * if ioengine defines a setup() method, it's responsible for
 735          * opening the files and setting f->real_file_size to indicate
 736          * the valid range for that file.
 737          */
 738         if (td->io_ops->setup)
 739                 err = td->io_ops->setup(td);
 740         else
 741                 err = get_file_sizes(td);
 742
 743         if (err)
 744                 return err;
 745
 746         /*
 747          * check sizes. if the files/devices do not exist and the size
 748          * isn't passed to fio, abort.
 749          */
 750         total_size = 0;
 751         for_each_file(td, f, i) {
 752                 if (f->real_file_size == -1ULL)
 753                         total_size = -1ULL;
 754                 else
 755                         total_size += f->real_file_size;
 756         }
 757
 758         if (td->o.fill_device)
 759                 td->fill_device_size = get_fs_free_counts(td);
 760
 761         /*
 762          * device/file sizes are zero and no size given, punt
 763          */
 764         if ((!total_size || total_size == -1ULL) && !td->o.size &&
 765             !(td->io_ops->flags & FIO_NOIO) && !td->o.fill_device) {
 766                 log_err("%s: you need to specify size=\n", td->o.name);
 767                 td_verror(td, EINVAL, "total_file_size");
 768                 return 1;
 769         }
 770
 771         /*
 772          * now file sizes are known, so we can set ->io_size. if size= is
 773          * not given, ->io_size is just equal to ->real_file_size. if size
 774          * is given, ->io_size is size / nr_files.
 775          */
 776         extend_size = total_size = 0;
 777         need_extend = 0;
 778         for_each_file(td, f, i) {
 779                 f->file_offset = get_start_offset(td);
 780
 781                 if (!td->o.file_size_low) {
 782                         /*
 783                          * no file size range given, file size is equal to
 784                          * total size divided by number of files. if that is
 785                          * zero, set it to the real file size.
 786                          */
 787                         f->io_size = td->o.size / td->o.nr_files;
 788                         if (!f->io_size)
 789                                 f->io_size = f->real_file_size - f->file_offset;
 790                 } else if (f->real_file_size < td->o.file_size_low ||
 791                            f->real_file_size > td->o.file_size_high) {
 792                         if (f->file_offset > td->o.file_size_low)
 793                                 goto err_offset;
 794                         /*
 795                          * file size given. if it's fixed, use that. if it's a
 796                          * range, generate a random size in-between.
 797                          */
 798                         if (td->o.file_size_low == td->o.file_size_high) {
 799                                 f->io_size = td->o.file_size_low
 800                                                 - f->file_offset;
 801                         } else {
 802                                 f->io_size = get_rand_file_size(td)
 803                                                 - f->file_offset;
 804                         }
 805                 } else
 806                         f->io_size = f->real_file_size - f->file_offset;
 807
 808                 if (f->io_size == -1ULL)
 809                         total_size = -1ULL;
 810                 else {
 811                         if (td->o.size_percent)
 812                                 f->io_size = (f->io_size * td->o.size_percent) / 100;
 813                         total_size += f->io_size;
 814                 }
 815
 816                 if (f->filetype == FIO_TYPE_FILE &&
 817                     (f->io_size + f->file_offset) > f->real_file_size &&
 818                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 819                         if (!td->o.create_on_open) {
 820                                 need_extend++;
 821                                 extend_size += (f->io_size + f->file_offset);
 822                         } else
 823                                 f->real_file_size = f->io_size + f->file_offset;
 824                         fio_file_set_extend(f);
 825                 }
 826         }
 827
 828         if (!td->o.size || td->o.size > total_size)
 829                 td->o.size = total_size;
 830
 831         /*
 832          * See if we need to extend some files
 833          */
 834         if (need_extend) {
 835                 temp_stall_ts = 1;
 836                 if (output_format == FIO_OUTPUT_NORMAL)
 837                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 838                                  " %lluMB)\n", td->o.name, need_extend,
 839                                         extend_size >> 20);
 840
 841                 for_each_file(td, f, i) {
 842                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 843
 844                         if (!fio_file_extend(f))
 845                                 continue;
 846
 847                         assert(f->filetype == FIO_TYPE_FILE);
 848                         fio_file_clear_extend(f);
 849                         if (!td->o.fill_device) {
 850                                 old_len = f->real_file_size;
 851                                 extend_len = f->io_size + f->file_offset -
 852                                                 old_len;
 853                         }
 854                         f->real_file_size = (f->io_size + f->file_offset);
 855                         err = extend_file(td, f);
 856                         if (err)
 857                                 break;
 858
 859                         err = __file_invalidate_cache(td, f, old_len,
 860                                                                 extend_len);
 861                         close(f->fd);
 862                         f->fd = -1;
 863                         if (err)
 864                                 break;
 865                 }
 866                 temp_stall_ts = 0;
 867         }
 868
 869         if (err)
 870                 return err;
 871
 872         if (!td->o.zone_size)
 873                 td->o.zone_size = td->o.size;
 874
 875         /*
 876          * iolog already set the total io size, if we read back
 877          * stored entries.
 878          */
 879         if (!td->o.read_iolog_file)
 880                 td->total_io_size = td->o.size * td->o.loops;
 881
 882 done:
 883         if (td->o.create_only)
 884                 td->done = 1;
 885
 886         return 0;
 887 err_offset:
 888         log_err("%s: you need to specify valid offset=\n", td->o.name);
 889         return 1;
 890 }
 891
 892 int pre_read_files(struct thread_data *td)
 893 {
 894         struct fio_file *f;
 895         unsigned int i;
 896
 897         dprint(FD_FILE, "pre_read files\n");
 898
 899         for_each_file(td, f, i) {
 900                 pre_read_file(td, f);
 901         }
 902
 903         return 1;
 904 }
 905
 906 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 907 {
 908         unsigned int range_size, seed;
 909         unsigned long nranges;
 910         uint64_t file_size;
 911
 912         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 913         file_size = min(f->real_file_size, f->io_size);
 914
 915         nranges = (file_size + range_size - 1) / range_size;
 916
 917         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 918         if (!td->o.rand_repeatable)
 919                 seed = td->rand_seeds[4];
 920
 921         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 922                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 923         else
 924                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 925
 926         return 1;
 927 }
 928
 929 static int init_rand_distribution(struct thread_data *td)
 930 {
 931         struct fio_file *f;
 932         unsigned int i;
 933         int state;
 934
 935         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 936                 return 0;
 937
 938         state = td->runstate;
 939         td_set_runstate(td, TD_SETTING_UP);
 940         for_each_file(td, f, i)
 941                 __init_rand_distribution(td, f);
 942         td_set_runstate(td, state);
 943
 944         return 1;
 945 }
 946
 947 int init_random_map(struct thread_data *td)
 948 {
 949         unsigned long long blocks;
 950         struct fio_file *f;
 951         unsigned int i;
 952
 953         if (init_rand_distribution(td))
 954                 return 0;
 955         if (!td_random(td))
 956                 return 0;
 957
 958         for_each_file(td, f, i) {
 959                 uint64_t file_size = min(f->real_file_size, f->io_size);
 960
 961                 blocks = (file_size + td->o.rw_min_bs - 1) /
 962                                 (unsigned long long) td->o.rw_min_bs;
 963                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
 964                         unsigned long seed;
 965
 966                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
 967
 968                         if (!lfsr_init(&f->lfsr, blocks, seed))
 969                                 continue;
 970                 } else if (!td->o.norandommap) {
 971                         f->io_axmap = axmap_new(blocks);
 972                         if (f->io_axmap)
 973                                 continue;
 974                 } else if (td->o.norandommap)
 975                         continue;
 976
 977                 if (!td->o.softrandommap) {
 978                         log_err("fio: failed allocating random map. If running"
 979                                 " a large number of jobs, try the 'norandommap'"
 980                                 " option or set 'softrandommap'. Or give"
 981                                 " a larger --alloc-size to fio.\n");
 982                         return 1;
 983                 }
 984
 985                 log_info("fio: file %s failed allocating random map. Running "
 986                          "job without.\n", f->file_name);
 987         }
 988
 989         return 0;
 990 }
 991
 992 void close_files(struct thread_data *td)
 993 {
 994         struct fio_file *f;
 995         unsigned int i;
 996
 997         for_each_file(td, f, i) {
 998                 if (fio_file_open(f))
 999                         td_io_close_file(td, f);
1000         }
1001 }
1002
1003 void close_and_free_files(struct thread_data *td)
1004 {
1005         struct fio_file *f;
1006         unsigned int i;
1007
1008         dprint(FD_FILE, "close files\n");
1009
1010         for_each_file(td, f, i) {
1011                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1012                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1013                         unlink(f->file_name);
1014                 }
1015
1016                 if (fio_file_open(f))
1017                         td_io_close_file(td, f);
1018
1019                 remove_file_hash(f);
1020
1021                 sfree(f->file_name);
1022                 f->file_name = NULL;
1023                 axmap_free(f->io_axmap);
1024                 f->io_axmap = NULL;
1025                 sfree(f);
1026         }
1027
1028         td->o.filename = NULL;
1029         free(td->files);
1030         td->files_index = 0;
1031         td->files = NULL;
1032         td->o.nr_files = 0;
1033 }
1034
1035 static void get_file_type(struct fio_file *f)
1036 {
1037         struct stat sb;
1038
1039         if (!strcmp(f->file_name, "-"))
1040                 f->filetype = FIO_TYPE_PIPE;
1041         else
1042                 f->filetype = FIO_TYPE_FILE;
1043
1044         /* \\.\ is the device namespace in Windows, where every file is
1045          * a block device */
1046         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1047                 f->filetype = FIO_TYPE_BD;
1048
1049         if (!stat(f->file_name, &sb)) {
1050                 if (S_ISBLK(sb.st_mode))
1051                         f->filetype = FIO_TYPE_BD;
1052                 else if (S_ISCHR(sb.st_mode))
1053                         f->filetype = FIO_TYPE_CHAR;
1054                 else if (S_ISFIFO(sb.st_mode))
1055                         f->filetype = FIO_TYPE_PIPE;
1056         }
1057 }
1058
1059 int add_file(struct thread_data *td, const char *fname)
1060 {
1061         int cur_files = td->files_index;
1062         char file_name[PATH_MAX];
1063         struct fio_file *f;
1064         int len = 0;
1065
1066         dprint(FD_FILE, "add file %s\n", fname);
1067
1068         f = smalloc(sizeof(*f));
1069         if (!f) {
1070                 log_err("fio: smalloc OOM\n");
1071                 assert(0);
1072         }
1073
1074         f->fd = -1;
1075         f->shadow_fd = -1;
1076         fio_file_reset(td, f);
1077
1078         if (td->files_size <= td->files_index) {
1079                 unsigned int new_size = td->o.nr_files + 1;
1080
1081                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1082
1083                 td->files = realloc(td->files, new_size * sizeof(f));
1084                 if (td->files == NULL) {
1085                         log_err("fio: realloc OOM\n");
1086                         assert(0);
1087                 }
1088                 td->files_size = new_size;
1089         }
1090         td->files[cur_files] = f;
1091         f->fileno = cur_files;
1092
1093         /*
1094          * init function, io engine may not be loaded yet
1095          */
1096         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1097                 f->real_file_size = -1ULL;
1098
1099         if (td->o.directory)
1100                 len = sprintf(file_name, "%s/", td->o.directory);
1101
1102         sprintf(file_name + len, "%s", fname);
1103         f->file_name = smalloc_strdup(file_name);
1104         if (!f->file_name) {
1105                 log_err("fio: smalloc OOM\n");
1106                 assert(0);
1107         }
1108
1109         get_file_type(f);
1110
1111         switch (td->o.file_lock_mode) {
1112         case FILE_LOCK_NONE:
1113                 break;
1114         case FILE_LOCK_READWRITE:
1115                 f->lock = fio_mutex_rw_init();
1116                 break;
1117         case FILE_LOCK_EXCLUSIVE:
1118                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1119                 break;
1120         default:
1121                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1122                 assert(0);
1123         }
1124
1125         td->files_index++;
1126         if (f->filetype == FIO_TYPE_FILE)
1127                 td->nr_normal_files++;
1128
1129         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1130                                                         cur_files);
1131
1132         return cur_files;
1133 }
1134
1135 int add_file_exclusive(struct thread_data *td, const char *fname)
1136 {
1137         struct fio_file *f;
1138         unsigned int i;
1139
1140         for_each_file(td, f, i) {
1141                 if (!strcmp(f->file_name, fname))
1142                         return i;
1143         }
1144
1145         return add_file(td, fname);
1146 }
1147
1148 void get_file(struct fio_file *f)
1149 {
1150         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1151         assert(fio_file_open(f));
1152         f->references++;
1153 }
1154
1155 int put_file(struct thread_data *td, struct fio_file *f)
1156 {
1157         int f_ret = 0, ret = 0;
1158
1159         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1160
1161         if (!fio_file_open(f)) {
1162                 assert(f->fd == -1);
1163                 return 0;
1164         }
1165
1166         assert(f->references);
1167         if (--f->references)
1168                 return 0;
1169
1170         if (should_fsync(td) && td->o.fsync_on_close)
1171                 f_ret = fsync(f->fd);
1172
1173         if (td->io_ops->close_file)
1174                 ret = td->io_ops->close_file(td, f);
1175
1176         if (!ret)
1177                 ret = f_ret;
1178
1179         td->nr_open_files--;
1180         fio_file_clear_open(f);
1181         assert(f->fd == -1);
1182         return ret;
1183 }
1184
1185 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1186 {
1187         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1188                 return;
1189
1190         if (f->lock_owner == td && f->lock_batch--)
1191                 return;
1192
1193         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1194                 if (ddir == DDIR_READ)
1195                         fio_mutex_down_read(f->lock);
1196                 else
1197                         fio_mutex_down_write(f->lock);
1198         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1199                 fio_mutex_down(f->lock);
1200
1201         f->lock_owner = td;
1202         f->lock_batch = td->o.lockfile_batch;
1203         f->lock_ddir = ddir;
1204 }
1205
1206 void unlock_file(struct thread_data *td, struct fio_file *f)
1207 {
1208         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1209                 return;
1210         if (f->lock_batch)
1211                 return;
1212
1213         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1214                 const int is_read = f->lock_ddir == DDIR_READ;
1215                 int val = fio_mutex_getval(f->lock);
1216
1217                 if ((is_read && val == 1) || (!is_read && val == -1))
1218                         f->lock_owner = NULL;
1219
1220                 if (is_read)
1221                         fio_mutex_up_read(f->lock);
1222                 else
1223                         fio_mutex_up_write(f->lock);
1224         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) {
1225                 int val = fio_mutex_getval(f->lock);
1226
1227                 if (val == 0)
1228                         f->lock_owner = NULL;
1229
1230                 fio_mutex_up(f->lock);
1231         }
1232 }
1233
1234 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1235 {
1236         if (f->lock_owner != td)
1237                 return;
1238
1239         f->lock_batch = 0;
1240         unlock_file(td, f);
1241 }
1242
1243 static int recurse_dir(struct thread_data *td, const char *dirname)
1244 {
1245         struct dirent *dir;
1246         int ret = 0;
1247         DIR *D;
1248
1249         D = opendir(dirname);
1250         if (!D) {
1251                 char buf[FIO_VERROR_SIZE];
1252
1253                 snprintf(buf, FIO_VERROR_SIZE - 1, "opendir(%s)", dirname);
1254                 td_verror(td, errno, buf);
1255                 return 1;
1256         }
1257
1258         while ((dir = readdir(D)) != NULL) {
1259                 char full_path[PATH_MAX];
1260                 struct stat sb;
1261
1262                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1263                         continue;
1264
1265                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1266
1267                 if (lstat(full_path, &sb) == -1) {
1268                         if (errno != ENOENT) {
1269                                 td_verror(td, errno, "stat");
1270                                 return 1;
1271                         }
1272                 }
1273
1274                 if (S_ISREG(sb.st_mode)) {
1275                         add_file(td, full_path);
1276                         td->o.nr_files++;
1277                         continue;
1278                 }
1279                 if (!S_ISDIR(sb.st_mode))
1280                         continue;
1281
1282                 ret = recurse_dir(td, full_path);
1283                 if (ret)
1284                         break;
1285         }
1286
1287         closedir(D);
1288         return ret;
1289 }
1290
1291 int add_dir_files(struct thread_data *td, const char *path)
1292 {
1293         int ret = recurse_dir(td, path);
1294
1295         if (!ret)
1296                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1297
1298         return ret;
1299 }
1300
1301 void dup_files(struct thread_data *td, struct thread_data *org)
1302 {
1303         struct fio_file *f;
1304         unsigned int i;
1305
1306         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1307
1308         if (!org->files)
1309                 return;
1310
1311         td->files = malloc(org->files_index * sizeof(f));
1312
1313         for_each_file(org, f, i) {
1314                 struct fio_file *__f;
1315
1316                 __f = smalloc(sizeof(*__f));
1317                 if (!__f) {
1318                         log_err("fio: smalloc OOM\n");
1319                         assert(0);
1320                 }
1321                 __f->fd = -1;
1322                 fio_file_reset(td, __f);
1323
1324                 if (f->file_name) {
1325                         __f->file_name = smalloc_strdup(f->file_name);
1326                         if (!__f->file_name) {
1327                                 log_err("fio: smalloc OOM\n");
1328                                 assert(0);
1329                         }
1330
1331                         __f->filetype = f->filetype;
1332                 }
1333
1334                 td->files[i] = __f;
1335         }
1336 }
1337
1338 /*
1339  * Returns the index that matches the filename, or -1 if not there
1340  */
1341 int get_fileno(struct thread_data *td, const char *fname)
1342 {
1343         struct fio_file *f;
1344         unsigned int i;
1345
1346         for_each_file(td, f, i)
1347                 if (!strcmp(f->file_name, fname))
1348                         return i;
1349
1350         return -1;
1351 }
1352
1353 /*
1354  * For log usage, where we add/open/close files automatically
1355  */
1356 void free_release_files(struct thread_data *td)
1357 {
1358         close_files(td);
1359         td->files_index = 0;
1360         td->nr_normal_files = 0;
1361 }
1362
1363 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1364 {
1365         f->last_pos = f->file_offset;
1366         f->last_start = -1ULL;
1367         if (f->io_axmap)
1368                 axmap_reset(f->io_axmap);
1369         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1370                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1371 }