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