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                 if (f->filetype == FIO_TYPE_BD)
 392                         (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 393 #endif
 394         } else if (f->filetype == FIO_TYPE_FILE) {
 395                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 396         } else if (f->filetype == FIO_TYPE_BD) {
 397                 ret = blockdev_invalidate_cache(f);
 398                 if (ret < 0 && errno == EACCES && geteuid()) {
 399                         if (!root_warn) {
 400                                 log_err("fio: only root may flush block "
 401                                         "devices. Cache flush bypassed!\n");
 402                                 root_warn = 1;
 403                         }
 404                         ret = 0;
 405                 }
 406         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 407                 ret = 0;
 408
 409         if (ret < 0) {
 410                 td_verror(td, errno, "invalidate_cache");
 411                 return 1;
 412         } else if (ret > 0) {
 413                 td_verror(td, ret, "invalidate_cache");
 414                 return 1;
 415         }
 416
 417         return ret;
 418
 419 }
 420
 421 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 422 {
 423         if (!fio_file_open(f))
 424                 return 0;
 425
 426         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 427 }
 428
 429 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 430 {
 431         int ret = 0;
 432
 433         dprint(FD_FILE, "fd close %s\n", f->file_name);
 434
 435         remove_file_hash(f);
 436
 437         if (close(f->fd) < 0)
 438                 ret = errno;
 439
 440         f->fd = -1;
 441
 442         if (f->shadow_fd != -1) {
 443                 close(f->shadow_fd);
 444                 f->shadow_fd = -1;
 445         }
 446
 447         f->engine_data = 0;
 448         return ret;
 449 }
 450
 451 int file_lookup_open(struct fio_file *f, int flags)
 452 {
 453         struct fio_file *__f;
 454         int from_hash;
 455
 456         __f = lookup_file_hash(f->file_name);
 457         if (__f) {
 458                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 459                 /*
 460                  * racy, need the __f->lock locked
 461                  */
 462                 f->lock = __f->lock;
 463                 from_hash = 1;
 464         } else {
 465                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 466                 from_hash = 0;
 467         }
 468
 469         f->fd = open(f->file_name, flags, 0600);
 470         return from_hash;
 471 }
 472
 473 static int file_close_shadow_fds(struct thread_data *td)
 474 {
 475         struct fio_file *f;
 476         int num_closed = 0;
 477         unsigned int i;
 478
 479         for_each_file(td, f, i) {
 480                 if (f->shadow_fd == -1)
 481                         continue;
 482
 483                 close(f->shadow_fd);
 484                 f->shadow_fd = -1;
 485                 num_closed++;
 486         }
 487
 488         return num_closed;
 489 }
 490
 491 int generic_open_file(struct thread_data *td, struct fio_file *f)
 492 {
 493         int is_std = 0;
 494         int flags = 0;
 495         int from_hash = 0;
 496
 497         dprint(FD_FILE, "fd open %s\n", f->file_name);
 498
 499         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 500                 log_err("fio: trim only applies to block device\n");
 501                 return 1;
 502         }
 503
 504         if (!strcmp(f->file_name, "-")) {
 505                 if (td_rw(td)) {
 506                         log_err("fio: can't read/write to stdin/out\n");
 507                         return 1;
 508                 }
 509                 is_std = 1;
 510
 511                 /*
 512                  * move output logging to stderr, if we are writing to stdout
 513                  */
 514                 if (td_write(td))
 515                         f_out = stderr;
 516         }
 517
 518         if (td_trim(td))
 519                 goto skip_flags;
 520         if (td->o.odirect)
 521                 flags |= OS_O_DIRECT;
 522         if (td->o.sync_io)
 523                 flags |= O_SYNC;
 524         if (td->o.create_on_open)
 525                 flags |= O_CREAT;
 526 skip_flags:
 527         if (f->filetype != FIO_TYPE_FILE)
 528                 flags |= FIO_O_NOATIME;
 529
 530 open_again:
 531         if (td_write(td)) {
 532                 if (!read_only)
 533                         flags |= O_RDWR;
 534
 535                 if (f->filetype == FIO_TYPE_FILE)
 536                         flags |= O_CREAT;
 537
 538                 if (is_std)
 539                         f->fd = dup(STDOUT_FILENO);
 540                 else
 541                         from_hash = file_lookup_open(f, flags);
 542         } else if (td_read(td)) {
 543                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 544                         flags |= O_RDWR;
 545                 else
 546                         flags |= O_RDONLY;
 547
 548                 if (is_std)
 549                         f->fd = dup(STDIN_FILENO);
 550                 else
 551                         from_hash = file_lookup_open(f, flags);
 552         } else { //td trim
 553                 flags |= O_RDWR;
 554                 from_hash = file_lookup_open(f, flags);
 555         }
 556
 557         if (f->fd == -1) {
 558                 char buf[FIO_VERROR_SIZE];
 559                 int __e = errno;
 560
 561                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 562                         flags &= ~FIO_O_NOATIME;
 563                         goto open_again;
 564                 }
 565                 if (__e == EMFILE && file_close_shadow_fds(td))
 566                         goto open_again;
 567
 568                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 569
 570                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 571                         log_err("fio: looks like your file system does not " \
 572                                 "support direct=1/buffered=0\n");
 573                 }
 574
 575                 td_verror(td, __e, buf);
 576         }
 577
 578         if (!from_hash && f->fd != -1) {
 579                 if (add_file_hash(f)) {
 580                         int fio_unused ret;
 581
 582                         /*
 583                          * Stash away descriptor for later close. This is to
 584                          * work-around a "feature" on Linux, where a close of
 585                          * an fd that has been opened for write will trigger
 586                          * udev to call blkid to check partitions, fs id, etc.
 587                          * That pollutes the device cache, which can slow down
 588                          * unbuffered accesses.
 589                          */
 590                         if (f->shadow_fd == -1)
 591                                 f->shadow_fd = f->fd;
 592                         else {
 593                                 /*
 594                                  * OK to ignore, we haven't done anything
 595                                  * with it
 596                                  */
 597                                 ret = generic_close_file(td, f);
 598                         }
 599                         goto open_again;
 600                 }
 601         }
 602
 603         return 0;
 604 }
 605
 606 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 607 {
 608         return get_file_size(td, f);
 609 }
 610
 611 /*
 612  * open/close all files, so that ->real_file_size gets set
 613  */
 614 static int get_file_sizes(struct thread_data *td)
 615 {
 616         struct fio_file *f;
 617         unsigned int i;
 618         int err = 0;
 619
 620         for_each_file(td, f, i) {
 621                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 622                                                                 f->file_name);
 623
 624                 if (td_io_get_file_size(td, f)) {
 625                         if (td->error != ENOENT) {
 626                                 log_err("%s\n", td->verror);
 627                                 err = 1;
 628                         }
 629                         clear_error(td);
 630                 }
 631
 632                 if (f->real_file_size == -1ULL && td->o.size)
 633                         f->real_file_size = td->o.size / td->o.nr_files;
 634         }
 635
 636         return err;
 637 }
 638
 639 struct fio_mount {
 640         struct flist_head list;
 641         const char *base;
 642         char __base[256];
 643         unsigned int key;
 644 };
 645
 646 /*
 647  * Get free number of bytes for each file on each unique mount.
 648  */
 649 static unsigned long long get_fs_free_counts(struct thread_data *td)
 650 {
 651         struct flist_head *n, *tmp;
 652         unsigned long long ret = 0;
 653         struct fio_mount *fm;
 654         FLIST_HEAD(list);
 655         struct fio_file *f;
 656         unsigned int i;
 657
 658         for_each_file(td, f, i) {
 659                 struct stat sb;
 660                 char buf[256];
 661
 662                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 663                         if (f->real_file_size != -1ULL)
 664                                 ret += f->real_file_size;
 665                         continue;
 666                 } else if (f->filetype != FIO_TYPE_FILE)
 667                         continue;
 668
 669                 strcpy(buf, f->file_name);
 670
 671                 if (stat(buf, &sb) < 0) {
 672                         if (errno != ENOENT)
 673                                 break;
 674                         strcpy(buf, ".");
 675                         if (stat(buf, &sb) < 0)
 676                                 break;
 677                 }
 678
 679                 fm = NULL;
 680                 flist_for_each(n, &list) {
 681                         fm = flist_entry(n, struct fio_mount, list);
 682                         if (fm->key == sb.st_dev)
 683                                 break;
 684
 685                         fm = NULL;
 686                 }
 687
 688                 if (fm)
 689                         continue;
 690
 691                 fm = malloc(sizeof(*fm));
 692                 strcpy(fm->__base, buf);
 693                 fm->base = basename(fm->__base);
 694                 fm->key = sb.st_dev;
 695                 flist_add(&fm->list, &list);
 696         }
 697
 698         flist_for_each_safe(n, tmp, &list) {
 699                 unsigned long long sz;
 700
 701                 fm = flist_entry(n, struct fio_mount, list);
 702                 flist_del(&fm->list);
 703
 704                 sz = get_fs_size(fm->base);
 705                 if (sz && sz != -1ULL)
 706                         ret += sz;
 707
 708                 free(fm);
 709         }
 710
 711         return ret;
 712 }
 713
 714 uint64_t get_start_offset(struct thread_data *td)
 715 {
 716         return td->o.start_offset +
 717                 (td->thread_number - 1) * td->o.offset_increment;
 718 }
 719
 720 /*
 721  * Open the files and setup files sizes, creating files if necessary.
 722  */
 723 int setup_files(struct thread_data *td)
 724 {
 725         unsigned long long total_size, extend_size;
 726         struct thread_options *o = &td->o;
 727         struct fio_file *f;
 728         unsigned int i;
 729         int err = 0, need_extend;
 730         int old_state;
 731
 732         dprint(FD_FILE, "setup files\n");
 733
 734         old_state = td->runstate;
 735         td_set_runstate(td, TD_SETTING_UP);
 736
 737         if (o->read_iolog_file)
 738                 goto done;
 739
 740         /*
 741          * if ioengine defines a setup() method, it's responsible for
 742          * opening the files and setting f->real_file_size to indicate
 743          * the valid range for that file.
 744          */
 745         if (td->io_ops->setup)
 746                 err = td->io_ops->setup(td);
 747         else
 748                 err = get_file_sizes(td);
 749
 750         if (err)
 751                 goto err_out;
 752
 753         /*
 754          * check sizes. if the files/devices do not exist and the size
 755          * isn't passed to fio, abort.
 756          */
 757         total_size = 0;
 758         for_each_file(td, f, i) {
 759                 if (f->real_file_size == -1ULL)
 760                         total_size = -1ULL;
 761                 else
 762                         total_size += f->real_file_size;
 763         }
 764
 765         if (o->fill_device)
 766                 td->fill_device_size = get_fs_free_counts(td);
 767
 768         /*
 769          * device/file sizes are zero and no size given, punt
 770          */
 771         if ((!total_size || total_size == -1ULL) && !o->size &&
 772             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 773             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 774                 log_err("%s: you need to specify size=\n", o->name);
 775                 td_verror(td, EINVAL, "total_file_size");
 776                 goto err_out;
 777         }
 778
 779         /*
 780          * now file sizes are known, so we can set ->io_size. if size= is
 781          * not given, ->io_size is just equal to ->real_file_size. if size
 782          * is given, ->io_size is size / nr_files.
 783          */
 784         extend_size = total_size = 0;
 785         need_extend = 0;
 786         for_each_file(td, f, i) {
 787                 f->file_offset = get_start_offset(td);
 788
 789                 if (!o->file_size_low) {
 790                         /*
 791                          * no file size range given, file size is equal to
 792                          * total size divided by number of files. if that is
 793                          * zero, set it to the real file size.
 794                          */
 795                         f->io_size = o->size / o->nr_files;
 796                         if (!f->io_size)
 797                                 f->io_size = f->real_file_size - f->file_offset;
 798                 } else if (f->real_file_size < o->file_size_low ||
 799                            f->real_file_size > o->file_size_high) {
 800                         if (f->file_offset > o->file_size_low)
 801                                 goto err_offset;
 802                         /*
 803                          * file size given. if it's fixed, use that. if it's a
 804                          * range, generate a random size in-between.
 805                          */
 806                         if (o->file_size_low == o->file_size_high)
 807                                 f->io_size = o->file_size_low - f->file_offset;
 808                         else {
 809                                 f->io_size = get_rand_file_size(td)
 810                                                 - f->file_offset;
 811                         }
 812                 } else
 813                         f->io_size = f->real_file_size - f->file_offset;
 814
 815                 if (f->io_size == -1ULL)
 816                         total_size = -1ULL;
 817                 else {
 818                         if (o->size_percent)
 819                                 f->io_size = (f->io_size * o->size_percent) / 100;
 820                         total_size += f->io_size;
 821                 }
 822
 823                 if (f->filetype == FIO_TYPE_FILE &&
 824                     (f->io_size + f->file_offset) > f->real_file_size &&
 825                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 826                         if (!o->create_on_open) {
 827                                 need_extend++;
 828                                 extend_size += (f->io_size + f->file_offset);
 829                         } else
 830                                 f->real_file_size = f->io_size + f->file_offset;
 831                         fio_file_set_extend(f);
 832                 }
 833         }
 834
 835         if (!o->size || o->size > total_size)
 836                 o->size = total_size;
 837
 838         /*
 839          * See if we need to extend some files
 840          */
 841         if (need_extend) {
 842                 temp_stall_ts = 1;
 843                 if (output_format == FIO_OUTPUT_NORMAL)
 844                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 845                                  " %lluMB)\n", o->name, need_extend,
 846                                         extend_size >> 20);
 847
 848                 for_each_file(td, f, i) {
 849                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 850
 851                         if (!fio_file_extend(f))
 852                                 continue;
 853
 854                         assert(f->filetype == FIO_TYPE_FILE);
 855                         fio_file_clear_extend(f);
 856                         if (!o->fill_device) {
 857                                 old_len = f->real_file_size;
 858                                 extend_len = f->io_size + f->file_offset -
 859                                                 old_len;
 860                         }
 861                         f->real_file_size = (f->io_size + f->file_offset);
 862                         err = extend_file(td, f);
 863                         if (err)
 864                                 break;
 865
 866                         err = __file_invalidate_cache(td, f, old_len,
 867                                                                 extend_len);
 868                         close(f->fd);
 869                         f->fd = -1;
 870                         if (err)
 871                                 break;
 872                 }
 873                 temp_stall_ts = 0;
 874         }
 875
 876         if (err)
 877                 goto err_out;
 878
 879         if (!o->zone_size)
 880                 o->zone_size = o->size;
 881
 882         /*
 883          * iolog already set the total io size, if we read back
 884          * stored entries.
 885          */
 886         if (!o->read_iolog_file)
 887                 td->total_io_size = o->size * o->loops;
 888
 889 done:
 890         if (o->create_only)
 891                 td->done = 1;
 892
 893         td_set_runstate(td, old_state);
 894         return 0;
 895 err_offset:
 896         log_err("%s: you need to specify valid offset=\n", o->name);
 897 err_out:
 898         td_set_runstate(td, old_state);
 899         return 1;
 900 }
 901
 902 int pre_read_files(struct thread_data *td)
 903 {
 904         struct fio_file *f;
 905         unsigned int i;
 906
 907         dprint(FD_FILE, "pre_read files\n");
 908
 909         for_each_file(td, f, i) {
 910                 pre_read_file(td, f);
 911         }
 912
 913         return 1;
 914 }
 915
 916 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 917 {
 918         unsigned int range_size, seed;
 919         unsigned long nranges;
 920         uint64_t file_size;
 921
 922         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 923         file_size = min(f->real_file_size, f->io_size);
 924
 925         nranges = (file_size + range_size - 1) / range_size;
 926
 927         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 928         if (!td->o.rand_repeatable)
 929                 seed = td->rand_seeds[4];
 930
 931         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 932                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 933         else
 934                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 935
 936         return 1;
 937 }
 938
 939 static int init_rand_distribution(struct thread_data *td)
 940 {
 941         struct fio_file *f;
 942         unsigned int i;
 943         int state;
 944
 945         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 946                 return 0;
 947
 948         state = td->runstate;
 949         td_set_runstate(td, TD_SETTING_UP);
 950         for_each_file(td, f, i)
 951                 __init_rand_distribution(td, f);
 952         td_set_runstate(td, state);
 953
 954         return 1;
 955 }
 956
 957 int init_random_map(struct thread_data *td)
 958 {
 959         unsigned long long blocks;
 960         struct fio_file *f;
 961         unsigned int i;
 962
 963         if (init_rand_distribution(td))
 964                 return 0;
 965         if (!td_random(td))
 966                 return 0;
 967
 968         for_each_file(td, f, i) {
 969                 uint64_t file_size = min(f->real_file_size, f->io_size);
 970
 971                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
 972
 973                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
 974                         unsigned long seed;
 975
 976                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
 977
 978                         if (!lfsr_init(&f->lfsr, blocks, seed, seed & 0xF))
 979                                 continue;
 980                 } else if (!td->o.norandommap) {
 981                         f->io_axmap = axmap_new(blocks);
 982                         if (f->io_axmap)
 983                                 continue;
 984                 } else if (td->o.norandommap)
 985                         continue;
 986
 987                 if (!td->o.softrandommap) {
 988                         log_err("fio: failed allocating random map. If running"
 989                                 " a large number of jobs, try the 'norandommap'"
 990                                 " option or set 'softrandommap'. Or give"
 991                                 " a larger --alloc-size to fio.\n");
 992                         return 1;
 993                 }
 994
 995                 log_info("fio: file %s failed allocating random map. Running "
 996                          "job without.\n", f->file_name);
 997         }
 998
 999         return 0;
1000 }
1001
1002 void close_files(struct thread_data *td)
1003 {
1004         struct fio_file *f;
1005         unsigned int i;
1006
1007         for_each_file(td, f, i) {
1008                 if (fio_file_open(f))
1009                         td_io_close_file(td, f);
1010         }
1011 }
1012
1013 void close_and_free_files(struct thread_data *td)
1014 {
1015         struct fio_file *f;
1016         unsigned int i;
1017
1018         dprint(FD_FILE, "close files\n");
1019
1020         for_each_file(td, f, i) {
1021                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1022                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1023                         unlink(f->file_name);
1024                 }
1025
1026                 if (fio_file_open(f))
1027                         td_io_close_file(td, f);
1028
1029                 remove_file_hash(f);
1030
1031                 sfree(f->file_name);
1032                 f->file_name = NULL;
1033                 axmap_free(f->io_axmap);
1034                 f->io_axmap = NULL;
1035                 sfree(f);
1036         }
1037
1038         td->o.filename = NULL;
1039         free(td->files);
1040         free(td->file_locks);
1041         td->files_index = 0;
1042         td->files = NULL;
1043         td->file_locks = NULL;
1044         td->o.nr_files = 0;
1045 }
1046
1047 static void get_file_type(struct fio_file *f)
1048 {
1049         struct stat sb;
1050
1051         if (!strcmp(f->file_name, "-"))
1052                 f->filetype = FIO_TYPE_PIPE;
1053         else
1054                 f->filetype = FIO_TYPE_FILE;
1055
1056         /* \\.\ is the device namespace in Windows, where every file is
1057          * a block device */
1058         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1059                 f->filetype = FIO_TYPE_BD;
1060
1061         if (!stat(f->file_name, &sb)) {
1062                 if (S_ISBLK(sb.st_mode))
1063                         f->filetype = FIO_TYPE_BD;
1064                 else if (S_ISCHR(sb.st_mode))
1065                         f->filetype = FIO_TYPE_CHAR;
1066                 else if (S_ISFIFO(sb.st_mode))
1067                         f->filetype = FIO_TYPE_PIPE;
1068         }
1069 }
1070
1071 int add_file(struct thread_data *td, const char *fname)
1072 {
1073         int cur_files = td->files_index;
1074         char file_name[PATH_MAX];
1075         struct fio_file *f;
1076         int len = 0;
1077
1078         dprint(FD_FILE, "add file %s\n", fname);
1079
1080         f = smalloc(sizeof(*f));
1081         if (!f) {
1082                 log_err("fio: smalloc OOM\n");
1083                 assert(0);
1084         }
1085
1086         f->fd = -1;
1087         f->shadow_fd = -1;
1088         fio_file_reset(td, f);
1089
1090         if (td->files_size <= td->files_index) {
1091                 unsigned int new_size = td->o.nr_files + 1;
1092
1093                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1094
1095                 td->files = realloc(td->files, new_size * sizeof(f));
1096                 if (td->files == NULL) {
1097                         log_err("fio: realloc OOM\n");
1098                         assert(0);
1099                 }
1100                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1101                         td->file_locks = realloc(td->file_locks, new_size);
1102                         if (!td->file_locks) {
1103                                 log_err("fio: realloc OOM\n");
1104                                 assert(0);
1105                         }
1106                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1107                 }
1108                 td->files_size = new_size;
1109         }
1110         td->files[cur_files] = f;
1111         f->fileno = cur_files;
1112
1113         /*
1114          * init function, io engine may not be loaded yet
1115          */
1116         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1117                 f->real_file_size = -1ULL;
1118
1119         if (td->o.directory)
1120                 len = sprintf(file_name, "%s/", td->o.directory);
1121
1122         sprintf(file_name + len, "%s", fname);
1123         f->file_name = smalloc_strdup(file_name);
1124         if (!f->file_name) {
1125                 log_err("fio: smalloc OOM\n");
1126                 assert(0);
1127         }
1128
1129         get_file_type(f);
1130
1131         switch (td->o.file_lock_mode) {
1132         case FILE_LOCK_NONE:
1133                 break;
1134         case FILE_LOCK_READWRITE:
1135                 f->rwlock = fio_rwlock_init();
1136                 break;
1137         case FILE_LOCK_EXCLUSIVE:
1138                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1139                 break;
1140         default:
1141                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1142                 assert(0);
1143         }
1144
1145         td->files_index++;
1146         if (f->filetype == FIO_TYPE_FILE)
1147                 td->nr_normal_files++;
1148
1149         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1150                                                         cur_files);
1151
1152         return cur_files;
1153 }
1154
1155 int add_file_exclusive(struct thread_data *td, const char *fname)
1156 {
1157         struct fio_file *f;
1158         unsigned int i;
1159
1160         for_each_file(td, f, i) {
1161                 if (!strcmp(f->file_name, fname))
1162                         return i;
1163         }
1164
1165         return add_file(td, fname);
1166 }
1167
1168 void get_file(struct fio_file *f)
1169 {
1170         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1171         assert(fio_file_open(f));
1172         f->references++;
1173 }
1174
1175 int put_file(struct thread_data *td, struct fio_file *f)
1176 {
1177         int f_ret = 0, ret = 0;
1178
1179         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1180
1181         if (!fio_file_open(f)) {
1182                 assert(f->fd == -1);
1183                 return 0;
1184         }
1185
1186         assert(f->references);
1187         if (--f->references)
1188                 return 0;
1189
1190         if (should_fsync(td) && td->o.fsync_on_close)
1191                 f_ret = fsync(f->fd);
1192
1193         if (td->io_ops->close_file)
1194                 ret = td->io_ops->close_file(td, f);
1195
1196         if (!ret)
1197                 ret = f_ret;
1198
1199         td->nr_open_files--;
1200         fio_file_clear_open(f);
1201         assert(f->fd == -1);
1202         return ret;
1203 }
1204
1205 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1206 {
1207         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1208                 return;
1209
1210         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1211                 if (ddir == DDIR_READ)
1212                         fio_rwlock_read(f->rwlock);
1213                 else
1214                         fio_rwlock_write(f->rwlock);
1215         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1216                 fio_mutex_down(f->lock);
1217
1218         td->file_locks[f->fileno] = td->o.file_lock_mode;
1219 }
1220
1221 void unlock_file(struct thread_data *td, struct fio_file *f)
1222 {
1223         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1224                 return;
1225
1226         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1227                 fio_rwlock_unlock(f->rwlock);
1228         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1229                 fio_mutex_up(f->lock);
1230
1231         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1232 }
1233
1234 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1235 {
1236         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1237                 unlock_file(td, f);
1238 }
1239
1240 static int recurse_dir(struct thread_data *td, const char *dirname)
1241 {
1242         struct dirent *dir;
1243         int ret = 0;
1244         DIR *D;
1245
1246         D = opendir(dirname);
1247         if (!D) {
1248                 char buf[FIO_VERROR_SIZE];
1249
1250                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1251                 td_verror(td, errno, buf);
1252                 return 1;
1253         }
1254
1255         while ((dir = readdir(D)) != NULL) {
1256                 char full_path[PATH_MAX];
1257                 struct stat sb;
1258
1259                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1260                         continue;
1261
1262                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1263
1264                 if (lstat(full_path, &sb) == -1) {
1265                         if (errno != ENOENT) {
1266                                 td_verror(td, errno, "stat");
1267                                 return 1;
1268                         }
1269                 }
1270
1271                 if (S_ISREG(sb.st_mode)) {
1272                         add_file(td, full_path);
1273                         td->o.nr_files++;
1274                         continue;
1275                 }
1276                 if (!S_ISDIR(sb.st_mode))
1277                         continue;
1278
1279                 ret = recurse_dir(td, full_path);
1280                 if (ret)
1281                         break;
1282         }
1283
1284         closedir(D);
1285         return ret;
1286 }
1287
1288 int add_dir_files(struct thread_data *td, const char *path)
1289 {
1290         int ret = recurse_dir(td, path);
1291
1292         if (!ret)
1293                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1294
1295         return ret;
1296 }
1297
1298 void dup_files(struct thread_data *td, struct thread_data *org)
1299 {
1300         struct fio_file *f;
1301         unsigned int i;
1302
1303         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1304
1305         if (!org->files)
1306                 return;
1307
1308         td->files = malloc(org->files_index * sizeof(f));
1309
1310         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1311                 td->file_locks = malloc(org->files_index);
1312
1313         for_each_file(org, f, i) {
1314                 struct fio_file *__f;
1315
1316                 __f = smalloc(sizeof(*__f));
1317                 if (!__f) {
1318                         log_err("fio: smalloc OOM\n");
1319                         assert(0);
1320                 }
1321                 __f->fd = -1;
1322                 fio_file_reset(td, __f);
1323
1324                 if (f->file_name) {
1325                         __f->file_name = smalloc_strdup(f->file_name);
1326                         if (!__f->file_name) {
1327                                 log_err("fio: smalloc OOM\n");
1328                                 assert(0);
1329                         }
1330
1331                         __f->filetype = f->filetype;
1332                 }
1333
1334                 td->files[i] = __f;
1335         }
1336 }
1337
1338 /*
1339  * Returns the index that matches the filename, or -1 if not there
1340  */
1341 int get_fileno(struct thread_data *td, const char *fname)
1342 {
1343         struct fio_file *f;
1344         unsigned int i;
1345
1346         for_each_file(td, f, i)
1347                 if (!strcmp(f->file_name, fname))
1348                         return i;
1349
1350         return -1;
1351 }
1352
1353 /*
1354  * For log usage, where we add/open/close files automatically
1355  */
1356 void free_release_files(struct thread_data *td)
1357 {
1358         close_files(td);
1359         td->files_index = 0;
1360         td->nr_normal_files = 0;
1361 }
1362
1363 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1364 {
1365         f->last_pos = f->file_offset;
1366         f->last_start = -1ULL;
1367         if (f->io_axmap)
1368                 axmap_reset(f->io_axmap);
1369         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1370                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1371 }