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
 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                 fill_io_buffer(td, b, bs, bs);
 138
 139                 r = write(f->fd, b, bs);
 140
 141                 if (r > 0) {
 142                         left -= r;
 143                         continue;
 144                 } else {
 145                         if (r < 0) {
 146                                 int __e = errno;
 147
 148                                 if (__e == ENOSPC) {
 149                                         if (td->o.fill_device)
 150                                                 break;
 151                                         log_info("fio: ENOSPC on laying out "
 152                                                  "file, stopping\n");
 153                                         break;
 154                                 }
 155                                 td_verror(td, errno, "write");
 156                         } else
 157                                 td_verror(td, EIO, "write");
 158
 159                         break;
 160                 }
 161         }
 162
 163         if (td->terminate) {
 164                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 165                 unlink(f->file_name);
 166         } else if (td->o.create_fsync) {
 167                 if (fsync(f->fd) < 0) {
 168                         td_verror(td, errno, "fsync");
 169                         goto err;
 170                 }
 171         }
 172         if (td->o.fill_device && !td_write(td)) {
 173                 fio_file_clear_size_known(f);
 174                 if (td_io_get_file_size(td, f))
 175                         goto err;
 176                 if (f->io_size > f->real_file_size)
 177                         f->io_size = f->real_file_size;
 178         }
 179
 180         free(b);
 181 done:
 182         return 0;
 183 err:
 184         close(f->fd);
 185         f->fd = -1;
 186         return 1;
 187 }
 188
 189 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 190 {
 191         int r, did_open = 0, old_runstate;
 192         unsigned long long left;
 193         unsigned int bs;
 194         char *b;
 195
 196         if (td->io_ops->flags & FIO_PIPEIO)
 197                 return 0;
 198
 199         if (!fio_file_open(f)) {
 200                 if (td->io_ops->open_file(td, f)) {
 201                         log_err("fio: cannot pre-read, failed to open file\n");
 202                         return 1;
 203                 }
 204                 did_open = 1;
 205         }
 206
 207         old_runstate = td->runstate;
 208         td_set_runstate(td, TD_PRE_READING);
 209
 210         bs = td->o.max_bs[DDIR_READ];
 211         b = malloc(bs);
 212         memset(b, 0, bs);
 213
 214         lseek(f->fd, f->file_offset, SEEK_SET);
 215         left = f->io_size;
 216
 217         while (left && !td->terminate) {
 218                 if (bs > left)
 219                         bs = left;
 220
 221                 r = read(f->fd, b, bs);
 222
 223                 if (r == (int) bs) {
 224                         left -= bs;
 225                         continue;
 226                 } else {
 227                         td_verror(td, EIO, "pre_read");
 228                         break;
 229                 }
 230         }
 231
 232         td_set_runstate(td, old_runstate);
 233
 234         if (did_open)
 235                 td->io_ops->close_file(td, f);
 236         free(b);
 237         return 0;
 238 }
 239
 240 static unsigned long long get_rand_file_size(struct thread_data *td)
 241 {
 242         unsigned long long ret, sized;
 243         unsigned long r;
 244
 245         if (td->o.use_os_rand) {
 246                 r = os_random_long(&td->file_size_state);
 247                 sized = td->o.file_size_high - td->o.file_size_low;
 248                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 249         } else {
 250                 r = __rand(&td->__file_size_state);
 251                 sized = td->o.file_size_high - td->o.file_size_low;
 252                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 253         }
 254
 255         ret += td->o.file_size_low;
 256         ret -= (ret % td->o.rw_min_bs);
 257         return ret;
 258 }
 259
 260 static int file_size(struct thread_data *td, struct fio_file *f)
 261 {
 262         struct stat st;
 263
 264         if (stat(f->file_name, &st) == -1) {
 265                 td_verror(td, errno, "fstat");
 266                 return 1;
 267         }
 268
 269         f->real_file_size = st.st_size;
 270         return 0;
 271 }
 272
 273 static int bdev_size(struct thread_data *td, struct fio_file *f)
 274 {
 275         unsigned long long bytes = 0;
 276         int r;
 277
 278         if (td->io_ops->open_file(td, f)) {
 279                 log_err("fio: failed opening blockdev %s for size check\n",
 280                         f->file_name);
 281                 return 1;
 282         }
 283
 284         r = blockdev_size(f, &bytes);
 285         if (r) {
 286                 td_verror(td, r, "blockdev_size");
 287                 goto err;
 288         }
 289
 290         if (!bytes) {
 291                 log_err("%s: zero sized block device?\n", f->file_name);
 292                 goto err;
 293         }
 294
 295         f->real_file_size = bytes;
 296         td->io_ops->close_file(td, f);
 297         return 0;
 298 err:
 299         td->io_ops->close_file(td, f);
 300         return 1;
 301 }
 302
 303 static int char_size(struct thread_data *td, struct fio_file *f)
 304 {
 305 #ifdef FIO_HAVE_CHARDEV_SIZE
 306         unsigned long long bytes = 0;
 307         int r;
 308
 309         if (td->io_ops->open_file(td, f)) {
 310                 log_err("fio: failed opening blockdev %s for size check\n",
 311                         f->file_name);
 312                 return 1;
 313         }
 314
 315         r = chardev_size(f, &bytes);
 316         if (r) {
 317                 td_verror(td, r, "chardev_size");
 318                 goto err;
 319         }
 320
 321         if (!bytes) {
 322                 log_err("%s: zero sized char device?\n", f->file_name);
 323                 goto err;
 324         }
 325
 326         f->real_file_size = bytes;
 327         td->io_ops->close_file(td, f);
 328         return 0;
 329 err:
 330         td->io_ops->close_file(td, f);
 331         return 1;
 332 #else
 333         f->real_file_size = -1ULL;
 334         return 0;
 335 #endif
 336 }
 337
 338 static int get_file_size(struct thread_data *td, struct fio_file *f)
 339 {
 340         int ret = 0;
 341
 342         if (fio_file_size_known(f))
 343                 return 0;
 344
 345         if (f->filetype == FIO_TYPE_FILE)
 346                 ret = file_size(td, f);
 347         else if (f->filetype == FIO_TYPE_BD)
 348                 ret = bdev_size(td, f);
 349         else if (f->filetype == FIO_TYPE_CHAR)
 350                 ret = char_size(td, f);
 351         else
 352                 f->real_file_size = -1;
 353
 354         if (ret)
 355                 return ret;
 356
 357         if (f->file_offset > f->real_file_size) {
 358                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 359                                         (unsigned long long) f->file_offset,
 360                                         (unsigned long long) f->real_file_size);
 361                 return 1;
 362         }
 363
 364         fio_file_set_size_known(f);
 365         return 0;
 366 }
 367
 368 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 369                                    unsigned long long off,
 370                                    unsigned long long len)
 371 {
 372         int ret = 0;
 373
 374         if (len == -1ULL)
 375                 len = f->io_size;
 376         if (off == -1ULL)
 377                 off = f->file_offset;
 378
 379         if (len == -1ULL || off == -1ULL)
 380                 return 0;
 381
 382         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 383                                                                 len);
 384
 385         /*
 386          * FIXME: add blockdev flushing too
 387          */
 388         if (f->mmap_ptr) {
 389                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 390 #ifdef FIO_MADV_FREE
 391                 (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 392 #endif
 393         } else if (f->filetype == FIO_TYPE_FILE) {
 394                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 395         } else if (f->filetype == FIO_TYPE_BD) {
 396                 ret = blockdev_invalidate_cache(f);
 397                 if (ret < 0 && errno == EACCES && geteuid()) {
 398                         if (!root_warn) {
 399                                 log_err("fio: only root may flush block "
 400                                         "devices. Cache flush bypassed!\n");
 401                                 root_warn = 1;
 402                         }
 403                         ret = 0;
 404                 }
 405         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 406                 ret = 0;
 407
 408         if (ret < 0) {
 409                 td_verror(td, errno, "invalidate_cache");
 410                 return 1;
 411         } else if (ret > 0) {
 412                 td_verror(td, ret, "invalidate_cache");
 413                 return 1;
 414         }
 415
 416         return ret;
 417
 418 }
 419
 420 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 421 {
 422         if (!fio_file_open(f))
 423                 return 0;
 424
 425         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 426 }
 427
 428 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 429 {
 430         int ret = 0;
 431
 432         dprint(FD_FILE, "fd close %s\n", f->file_name);
 433
 434         remove_file_hash(f);
 435
 436         if (close(f->fd) < 0)
 437                 ret = errno;
 438
 439         f->fd = -1;
 440
 441         if (f->shadow_fd != -1) {
 442                 close(f->shadow_fd);
 443                 f->shadow_fd = -1;
 444         }
 445
 446         return ret;
 447 }
 448
 449 int file_lookup_open(struct fio_file *f, int flags)
 450 {
 451         struct fio_file *__f;
 452         int from_hash;
 453
 454         __f = lookup_file_hash(f->file_name);
 455         if (__f) {
 456                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 457                 /*
 458                  * racy, need the __f->lock locked
 459                  */
 460                 f->lock = __f->lock;
 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), "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 thread_options *o = &td->o;
 725         struct fio_file *f;
 726         unsigned int i;
 727         int err = 0, need_extend;
 728         int old_state;
 729
 730         dprint(FD_FILE, "setup files\n");
 731
 732         old_state = td->runstate;
 733         td_set_runstate(td, TD_SETTING_UP);
 734
 735         if (o->read_iolog_file)
 736                 goto done;
 737
 738         /*
 739          * if ioengine defines a setup() method, it's responsible for
 740          * opening the files and setting f->real_file_size to indicate
 741          * the valid range for that file.
 742          */
 743         if (td->io_ops->setup)
 744                 err = td->io_ops->setup(td);
 745         else
 746                 err = get_file_sizes(td);
 747
 748         if (err)
 749                 goto err_out;
 750
 751         /*
 752          * check sizes. if the files/devices do not exist and the size
 753          * isn't passed to fio, abort.
 754          */
 755         total_size = 0;
 756         for_each_file(td, f, i) {
 757                 if (f->real_file_size == -1ULL)
 758                         total_size = -1ULL;
 759                 else
 760                         total_size += f->real_file_size;
 761         }
 762
 763         if (o->fill_device)
 764                 td->fill_device_size = get_fs_free_counts(td);
 765
 766         /*
 767          * device/file sizes are zero and no size given, punt
 768          */
 769         if ((!total_size || total_size == -1ULL) && !o->size &&
 770             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 771             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 772                 log_err("%s: you need to specify size=\n", o->name);
 773                 td_verror(td, EINVAL, "total_file_size");
 774                 goto err_out;
 775         }
 776
 777         /*
 778          * now file sizes are known, so we can set ->io_size. if size= is
 779          * not given, ->io_size is just equal to ->real_file_size. if size
 780          * is given, ->io_size is size / nr_files.
 781          */
 782         extend_size = total_size = 0;
 783         need_extend = 0;
 784         for_each_file(td, f, i) {
 785                 f->file_offset = get_start_offset(td);
 786
 787                 if (!o->file_size_low) {
 788                         /*
 789                          * no file size range given, file size is equal to
 790                          * total size divided by number of files. if that is
 791                          * zero, set it to the real file size.
 792                          */
 793                         f->io_size = o->size / o->nr_files;
 794                         if (!f->io_size)
 795                                 f->io_size = f->real_file_size - f->file_offset;
 796                 } else if (f->real_file_size < o->file_size_low ||
 797                            f->real_file_size > o->file_size_high) {
 798                         if (f->file_offset > o->file_size_low)
 799                                 goto err_offset;
 800                         /*
 801                          * file size given. if it's fixed, use that. if it's a
 802                          * range, generate a random size in-between.
 803                          */
 804                         if (o->file_size_low == o->file_size_high)
 805                                 f->io_size = o->file_size_low - f->file_offset;
 806                         else {
 807                                 f->io_size = get_rand_file_size(td)
 808                                                 - f->file_offset;
 809                         }
 810                 } else
 811                         f->io_size = f->real_file_size - f->file_offset;
 812
 813                 if (f->io_size == -1ULL)
 814                         total_size = -1ULL;
 815                 else {
 816                         if (o->size_percent)
 817                                 f->io_size = (f->io_size * o->size_percent) / 100;
 818                         total_size += f->io_size;
 819                 }
 820
 821                 if (f->filetype == FIO_TYPE_FILE &&
 822                     (f->io_size + f->file_offset) > f->real_file_size &&
 823                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 824                         if (!o->create_on_open) {
 825                                 need_extend++;
 826                                 extend_size += (f->io_size + f->file_offset);
 827                         } else
 828                                 f->real_file_size = f->io_size + f->file_offset;
 829                         fio_file_set_extend(f);
 830                 }
 831         }
 832
 833         if (!o->size || o->size > total_size)
 834                 o->size = total_size;
 835
 836         /*
 837          * See if we need to extend some files
 838          */
 839         if (need_extend) {
 840                 temp_stall_ts = 1;
 841                 if (output_format == FIO_OUTPUT_NORMAL)
 842                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 843                                  " %lluMB)\n", o->name, need_extend,
 844                                         extend_size >> 20);
 845
 846                 for_each_file(td, f, i) {
 847                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 848
 849                         if (!fio_file_extend(f))
 850                                 continue;
 851
 852                         assert(f->filetype == FIO_TYPE_FILE);
 853                         fio_file_clear_extend(f);
 854                         if (!o->fill_device) {
 855                                 old_len = f->real_file_size;
 856                                 extend_len = f->io_size + f->file_offset -
 857                                                 old_len;
 858                         }
 859                         f->real_file_size = (f->io_size + f->file_offset);
 860                         err = extend_file(td, f);
 861                         if (err)
 862                                 break;
 863
 864                         err = __file_invalidate_cache(td, f, old_len,
 865                                                                 extend_len);
 866                         close(f->fd);
 867                         f->fd = -1;
 868                         if (err)
 869                                 break;
 870                 }
 871                 temp_stall_ts = 0;
 872         }
 873
 874         if (err)
 875                 goto err_out;
 876
 877         if (!o->zone_size)
 878                 o->zone_size = o->size;
 879
 880         /*
 881          * iolog already set the total io size, if we read back
 882          * stored entries.
 883          */
 884         if (!o->read_iolog_file)
 885                 td->total_io_size = o->size * o->loops;
 886
 887 done:
 888         if (o->create_only)
 889                 td->done = 1;
 890
 891         td_set_runstate(td, old_state);
 892         return 0;
 893 err_offset:
 894         log_err("%s: you need to specify valid offset=\n", o->name);
 895 err_out:
 896         td_set_runstate(td, old_state);
 897         return 1;
 898 }
 899
 900 int pre_read_files(struct thread_data *td)
 901 {
 902         struct fio_file *f;
 903         unsigned int i;
 904
 905         dprint(FD_FILE, "pre_read files\n");
 906
 907         for_each_file(td, f, i) {
 908                 pre_read_file(td, f);
 909         }
 910
 911         return 1;
 912 }
 913
 914 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 915 {
 916         unsigned int range_size, seed;
 917         unsigned long nranges;
 918         uint64_t file_size;
 919
 920         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 921         file_size = min(f->real_file_size, f->io_size);
 922
 923         nranges = (file_size + range_size - 1) / range_size;
 924
 925         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 926         if (!td->o.rand_repeatable)
 927                 seed = td->rand_seeds[4];
 928
 929         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 930                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 931         else
 932                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 933
 934         return 1;
 935 }
 936
 937 static int init_rand_distribution(struct thread_data *td)
 938 {
 939         struct fio_file *f;
 940         unsigned int i;
 941         int state;
 942
 943         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 944                 return 0;
 945
 946         state = td->runstate;
 947         td_set_runstate(td, TD_SETTING_UP);
 948         for_each_file(td, f, i)
 949                 __init_rand_distribution(td, f);
 950         td_set_runstate(td, state);
 951
 952         return 1;
 953 }
 954
 955 int init_random_map(struct thread_data *td)
 956 {
 957         unsigned long long blocks;
 958         struct fio_file *f;
 959         unsigned int i;
 960
 961         if (init_rand_distribution(td))
 962                 return 0;
 963         if (!td_random(td))
 964                 return 0;
 965
 966         for_each_file(td, f, i) {
 967                 uint64_t file_size = min(f->real_file_size, f->io_size);
 968
 969                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
 970
 971                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
 972                         unsigned long seed;
 973
 974                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
 975
 976                         if (!lfsr_init(&f->lfsr, blocks, seed, seed & 0xF))
 977                                 continue;
 978                 } else if (!td->o.norandommap) {
 979                         f->io_axmap = axmap_new(blocks);
 980                         if (f->io_axmap)
 981                                 continue;
 982                 } else if (td->o.norandommap)
 983                         continue;
 984
 985                 if (!td->o.softrandommap) {
 986                         log_err("fio: failed allocating random map. If running"
 987                                 " a large number of jobs, try the 'norandommap'"
 988                                 " option or set 'softrandommap'. Or give"
 989                                 " a larger --alloc-size to fio.\n");
 990                         return 1;
 991                 }
 992
 993                 log_info("fio: file %s failed allocating random map. Running "
 994                          "job without.\n", f->file_name);
 995         }
 996
 997         return 0;
 998 }
 999
1000 void close_files(struct thread_data *td)
1001 {
1002         struct fio_file *f;
1003         unsigned int i;
1004
1005         for_each_file(td, f, i) {
1006                 if (fio_file_open(f))
1007                         td_io_close_file(td, f);
1008         }
1009 }
1010
1011 void close_and_free_files(struct thread_data *td)
1012 {
1013         struct fio_file *f;
1014         unsigned int i;
1015
1016         dprint(FD_FILE, "close files\n");
1017
1018         for_each_file(td, f, i) {
1019                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1020                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1021                         unlink(f->file_name);
1022                 }
1023
1024                 if (fio_file_open(f))
1025                         td_io_close_file(td, f);
1026
1027                 remove_file_hash(f);
1028
1029                 sfree(f->file_name);
1030                 f->file_name = NULL;
1031                 axmap_free(f->io_axmap);
1032                 f->io_axmap = NULL;
1033                 sfree(f);
1034         }
1035
1036         td->o.filename = NULL;
1037         free(td->files);
1038         free(td->file_locks);
1039         td->files_index = 0;
1040         td->files = NULL;
1041         td->file_locks = NULL;
1042         td->o.nr_files = 0;
1043 }
1044
1045 static void get_file_type(struct fio_file *f)
1046 {
1047         struct stat sb;
1048
1049         if (!strcmp(f->file_name, "-"))
1050                 f->filetype = FIO_TYPE_PIPE;
1051         else
1052                 f->filetype = FIO_TYPE_FILE;
1053
1054         /* \\.\ is the device namespace in Windows, where every file is
1055          * a block device */
1056         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1057                 f->filetype = FIO_TYPE_BD;
1058
1059         if (!stat(f->file_name, &sb)) {
1060                 if (S_ISBLK(sb.st_mode))
1061                         f->filetype = FIO_TYPE_BD;
1062                 else if (S_ISCHR(sb.st_mode))
1063                         f->filetype = FIO_TYPE_CHAR;
1064                 else if (S_ISFIFO(sb.st_mode))
1065                         f->filetype = FIO_TYPE_PIPE;
1066         }
1067 }
1068
1069 int add_file(struct thread_data *td, const char *fname)
1070 {
1071         int cur_files = td->files_index;
1072         char file_name[PATH_MAX];
1073         struct fio_file *f;
1074         int len = 0;
1075
1076         dprint(FD_FILE, "add file %s\n", fname);
1077
1078         f = smalloc(sizeof(*f));
1079         if (!f) {
1080                 log_err("fio: smalloc OOM\n");
1081                 assert(0);
1082         }
1083
1084         f->fd = -1;
1085         f->shadow_fd = -1;
1086         fio_file_reset(td, f);
1087
1088         if (td->files_size <= td->files_index) {
1089                 unsigned int new_size = td->o.nr_files + 1;
1090
1091                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1092
1093                 td->files = realloc(td->files, new_size * sizeof(f));
1094                 if (td->files == NULL) {
1095                         log_err("fio: realloc OOM\n");
1096                         assert(0);
1097                 }
1098                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1099                         td->file_locks = realloc(td->file_locks, new_size);
1100                         if (!td->file_locks) {
1101                                 log_err("fio: realloc OOM\n");
1102                                 assert(0);
1103                         }
1104                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1105                 }
1106                 td->files_size = new_size;
1107         }
1108         td->files[cur_files] = f;
1109         f->fileno = cur_files;
1110
1111         /*
1112          * init function, io engine may not be loaded yet
1113          */
1114         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1115                 f->real_file_size = -1ULL;
1116
1117         if (td->o.directory)
1118                 len = sprintf(file_name, "%s/", td->o.directory);
1119
1120         sprintf(file_name + len, "%s", fname);
1121         f->file_name = smalloc_strdup(file_name);
1122         if (!f->file_name) {
1123                 log_err("fio: smalloc OOM\n");
1124                 assert(0);
1125         }
1126
1127         get_file_type(f);
1128
1129         switch (td->o.file_lock_mode) {
1130         case FILE_LOCK_NONE:
1131                 break;
1132         case FILE_LOCK_READWRITE:
1133                 f->rwlock = fio_rwlock_init();
1134                 break;
1135         case FILE_LOCK_EXCLUSIVE:
1136                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1137                 break;
1138         default:
1139                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1140                 assert(0);
1141         }
1142
1143         td->files_index++;
1144         if (f->filetype == FIO_TYPE_FILE)
1145                 td->nr_normal_files++;
1146
1147         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1148                                                         cur_files);
1149
1150         return cur_files;
1151 }
1152
1153 int add_file_exclusive(struct thread_data *td, const char *fname)
1154 {
1155         struct fio_file *f;
1156         unsigned int i;
1157
1158         for_each_file(td, f, i) {
1159                 if (!strcmp(f->file_name, fname))
1160                         return i;
1161         }
1162
1163         return add_file(td, fname);
1164 }
1165
1166 void get_file(struct fio_file *f)
1167 {
1168         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1169         assert(fio_file_open(f));
1170         f->references++;
1171 }
1172
1173 int put_file(struct thread_data *td, struct fio_file *f)
1174 {
1175         int f_ret = 0, ret = 0;
1176
1177         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1178
1179         if (!fio_file_open(f)) {
1180                 assert(f->fd == -1);
1181                 return 0;
1182         }
1183
1184         assert(f->references);
1185         if (--f->references)
1186                 return 0;
1187
1188         if (should_fsync(td) && td->o.fsync_on_close)
1189                 f_ret = fsync(f->fd);
1190
1191         if (td->io_ops->close_file)
1192                 ret = td->io_ops->close_file(td, f);
1193
1194         if (!ret)
1195                 ret = f_ret;
1196
1197         td->nr_open_files--;
1198         fio_file_clear_open(f);
1199         assert(f->fd == -1);
1200         return ret;
1201 }
1202
1203 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1204 {
1205         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1206                 return;
1207
1208         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1209                 if (ddir == DDIR_READ)
1210                         fio_rwlock_read(f->rwlock);
1211                 else
1212                         fio_rwlock_write(f->rwlock);
1213         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1214                 fio_mutex_down(f->lock);
1215
1216         td->file_locks[f->fileno] = td->o.file_lock_mode;
1217 }
1218
1219 void unlock_file(struct thread_data *td, struct fio_file *f)
1220 {
1221         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1222                 return;
1223
1224         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1225                 fio_rwlock_unlock(f->rwlock);
1226         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1227                 fio_mutex_up(f->lock);
1228
1229         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1230 }
1231
1232 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1233 {
1234         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1235                 unlock_file(td, f);
1236 }
1237
1238 static int recurse_dir(struct thread_data *td, const char *dirname)
1239 {
1240         struct dirent *dir;
1241         int ret = 0;
1242         DIR *D;
1243
1244         D = opendir(dirname);
1245         if (!D) {
1246                 char buf[FIO_VERROR_SIZE];
1247
1248                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1249                 td_verror(td, errno, buf);
1250                 return 1;
1251         }
1252
1253         while ((dir = readdir(D)) != NULL) {
1254                 char full_path[PATH_MAX];
1255                 struct stat sb;
1256
1257                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1258                         continue;
1259
1260                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1261
1262                 if (lstat(full_path, &sb) == -1) {
1263                         if (errno != ENOENT) {
1264                                 td_verror(td, errno, "stat");
1265                                 return 1;
1266                         }
1267                 }
1268
1269                 if (S_ISREG(sb.st_mode)) {
1270                         add_file(td, full_path);
1271                         td->o.nr_files++;
1272                         continue;
1273                 }
1274                 if (!S_ISDIR(sb.st_mode))
1275                         continue;
1276
1277                 ret = recurse_dir(td, full_path);
1278                 if (ret)
1279                         break;
1280         }
1281
1282         closedir(D);
1283         return ret;
1284 }
1285
1286 int add_dir_files(struct thread_data *td, const char *path)
1287 {
1288         int ret = recurse_dir(td, path);
1289
1290         if (!ret)
1291                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1292
1293         return ret;
1294 }
1295
1296 void dup_files(struct thread_data *td, struct thread_data *org)
1297 {
1298         struct fio_file *f;
1299         unsigned int i;
1300
1301         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1302
1303         if (!org->files)
1304                 return;
1305
1306         td->files = malloc(org->files_index * sizeof(f));
1307
1308         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1309                 td->file_locks = malloc(org->files_index);
1310
1311         for_each_file(org, f, i) {
1312                 struct fio_file *__f;
1313
1314                 __f = smalloc(sizeof(*__f));
1315                 if (!__f) {
1316                         log_err("fio: smalloc OOM\n");
1317                         assert(0);
1318                 }
1319                 __f->fd = -1;
1320                 fio_file_reset(td, __f);
1321
1322                 if (f->file_name) {
1323                         __f->file_name = smalloc_strdup(f->file_name);
1324                         if (!__f->file_name) {
1325                                 log_err("fio: smalloc OOM\n");
1326                                 assert(0);
1327                         }
1328
1329                         __f->filetype = f->filetype;
1330                 }
1331
1332                 td->files[i] = __f;
1333         }
1334 }
1335
1336 /*
1337  * Returns the index that matches the filename, or -1 if not there
1338  */
1339 int get_fileno(struct thread_data *td, const char *fname)
1340 {
1341         struct fio_file *f;
1342         unsigned int i;
1343
1344         for_each_file(td, f, i)
1345                 if (!strcmp(f->file_name, fname))
1346                         return i;
1347
1348         return -1;
1349 }
1350
1351 /*
1352  * For log usage, where we add/open/close files automatically
1353  */
1354 void free_release_files(struct thread_data *td)
1355 {
1356         close_files(td);
1357         td->files_index = 0;
1358         td->nr_normal_files = 0;
1359 }
1360
1361 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1362 {
1363         f->last_pos = f->file_offset;
1364         f->last_start = -1ULL;
1365         if (f->io_axmap)
1366                 axmap_reset(f->io_axmap);
1367         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1368                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1369 }