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