filesetup.c

   1 #include <unistd.h>
   2 #include <fcntl.h>
   3 #include <string.h>
   4 #include <assert.h>
   5 #include <dirent.h>
   6 #include <libgen.h>
   7 #include <sys/stat.h>
   8 #include <sys/mman.h>
   9 #include <sys/types.h>
  10
  11 #include "fio.h"
  12 #include "smalloc.h"
  13 #include "filehash.h"
  14 #include "os/os.h"
  15 #include "hash.h"
  16 #include "lib/axmap.h"
  17
  18 #ifdef CONFIG_LINUX_FALLOCATE
  19 #include <linux/falloc.h>
  20 #endif
  21
  22 static int root_warn;
  23
  24 static inline void clear_error(struct thread_data *td)
  25 {
  26         td->error = 0;
  27         td->verror[0] = '\0';
  28 }
  29
  30 /*
  31  * Leaves f->fd open on success, caller must close
  32  */
  33 static int extend_file(struct thread_data *td, struct fio_file *f)
  34 {
  35         int r, new_layout = 0, unlink_file = 0, flags;
  36         unsigned long long left;
  37         unsigned int bs;
  38         char *b;
  39
  40         if (read_only) {
  41                 log_err("fio: refusing extend of file due to read-only\n");
  42                 return 0;
  43         }
  44
  45         /*
  46          * check if we need to lay the file out complete again. fio
  47          * does that for operations involving reads, or for writes
  48          * where overwrite is set
  49          */
  50         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
  51             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
  52                 new_layout = 1;
  53         if (td_write(td) && !td->o.overwrite)
  54                 unlink_file = 1;
  55
  56         if (unlink_file || new_layout) {
  57                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  58                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
  59                         td_verror(td, errno, "unlink");
  60                         return 1;
  61                 }
  62         }
  63
  64         flags = O_WRONLY | O_CREAT;
  65         if (new_layout)
  66                 flags |= O_TRUNC;
  67
  68         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  69         f->fd = open(f->file_name, flags, 0644);
  70         if (f->fd < 0) {
  71                 td_verror(td, errno, "open");
  72                 return 1;
  73         }
  74
  75 #ifdef CONFIG_POSIX_FALLOCATE
  76         if (!td->o.fill_device) {
  77                 switch (td->o.fallocate_mode) {
  78                 case FIO_FALLOCATE_NONE:
  79                         break;
  80                 case FIO_FALLOCATE_POSIX:
  81                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  82                                  f->file_name,
  83                                  (unsigned long long) f->real_file_size);
  84
  85                         r = posix_fallocate(f->fd, 0, f->real_file_size);
  86                         if (r > 0) {
  87                                 log_err("fio: posix_fallocate fails: %s\n",
  88                                                 strerror(r));
  89                         }
  90                         break;
  91 #ifdef CONFIG_LINUX_FALLOCATE
  92                 case FIO_FALLOCATE_KEEP_SIZE:
  93                         dprint(FD_FILE,
  94                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
  95                                 "file %s size %llu\n", f->file_name,
  96                                 (unsigned long long) f->real_file_size);
  97
  98                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
  99                                         f->real_file_size);
 100                         if (r != 0) {
 101                                 td_verror(td, errno, "fallocate");
 102                         }
 103                         break;
 104 #endif /* CONFIG_LINUX_FALLOCATE */
 105                 default:
 106                         log_err("fio: unknown fallocate mode: %d\n",
 107                                 td->o.fallocate_mode);
 108                         assert(0);
 109                 }
 110         }
 111 #endif /* CONFIG_POSIX_FALLOCATE */
 112
 113         if (!new_layout)
 114                 goto done;
 115
 116         /*
 117          * The size will be -1ULL when fill_device is used, so don't truncate
 118          * or fallocate this file, just write it
 119          */
 120         if (!td->o.fill_device) {
 121                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 122                                         (unsigned long long) f->real_file_size);
 123                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 124                         td_verror(td, errno, "ftruncate");
 125                         goto err;
 126                 }
 127         }
 128
 129         b = malloc(td->o.max_bs[DDIR_WRITE]);
 130         memset(b, 0, td->o.max_bs[DDIR_WRITE]);
 131
 132         left = f->real_file_size;
 133         while (left && !td->terminate) {
 134                 bs = td->o.max_bs[DDIR_WRITE];
 135                 if (bs > left)
 136                         bs = left;
 137
 138                 r = write(f->fd, b, bs);
 139
 140                 if (r > 0) {
 141                         left -= r;
 142                         continue;
 143                 } else {
 144                         if (r < 0) {
 145                                 int __e = errno;
 146
 147                                 if (__e == ENOSPC) {
 148                                         if (td->o.fill_device)
 149                                                 break;
 150                                         log_info("fio: ENOSPC on laying out "
 151                                                  "file, stopping\n");
 152                                         break;
 153                                 }
 154                                 td_verror(td, errno, "write");
 155                         } else
 156                                 td_verror(td, EIO, "write");
 157
 158                         break;
 159                 }
 160         }
 161
 162         if (td->terminate) {
 163                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 164                 unlink(f->file_name);
 165         } else if (td->o.create_fsync) {
 166                 if (fsync(f->fd) < 0) {
 167                         td_verror(td, errno, "fsync");
 168                         goto err;
 169                 }
 170         }
 171         if (td->o.fill_device && !td_write(td)) {
 172                 fio_file_clear_size_known(f);
 173                 if (td_io_get_file_size(td, f))
 174                         goto err;
 175                 if (f->io_size > f->real_file_size)
 176                         f->io_size = f->real_file_size;
 177         }
 178
 179         free(b);
 180 done:
 181         return 0;
 182 err:
 183         close(f->fd);
 184         f->fd = -1;
 185         return 1;
 186 }
 187
 188 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 189 {
 190         int r, did_open = 0, old_runstate;
 191         unsigned long long left;
 192         unsigned int bs;
 193         char *b;
 194
 195         if (td->io_ops->flags & FIO_PIPEIO)
 196                 return 0;
 197
 198         if (!fio_file_open(f)) {
 199                 if (td->io_ops->open_file(td, f)) {
 200                         log_err("fio: cannot pre-read, failed to open file\n");
 201                         return 1;
 202                 }
 203                 did_open = 1;
 204         }
 205
 206         old_runstate = td->runstate;
 207         td_set_runstate(td, TD_PRE_READING);
 208
 209         bs = td->o.max_bs[DDIR_READ];
 210         b = malloc(bs);
 211         memset(b, 0, bs);
 212
 213         lseek(f->fd, f->file_offset, SEEK_SET);
 214         left = f->io_size;
 215
 216         while (left && !td->terminate) {
 217                 if (bs > left)
 218                         bs = left;
 219
 220                 r = read(f->fd, b, bs);
 221
 222                 if (r == (int) bs) {
 223                         left -= bs;
 224                         continue;
 225                 } else {
 226                         td_verror(td, EIO, "pre_read");
 227                         break;
 228                 }
 229         }
 230
 231         td_set_runstate(td, old_runstate);
 232
 233         if (did_open)
 234                 td->io_ops->close_file(td, f);
 235         free(b);
 236         return 0;
 237 }
 238
 239 static unsigned long long get_rand_file_size(struct thread_data *td)
 240 {
 241         unsigned long long ret, sized;
 242         unsigned long r;
 243
 244         if (td->o.use_os_rand) {
 245                 r = os_random_long(&td->file_size_state);
 246                 sized = td->o.file_size_high - td->o.file_size_low;
 247                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 248         } else {
 249                 r = __rand(&td->__file_size_state);
 250                 sized = td->o.file_size_high - td->o.file_size_low;
 251                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 252         }
 253
 254         ret += td->o.file_size_low;
 255         ret -= (ret % td->o.rw_min_bs);
 256         return ret;
 257 }
 258
 259 static int file_size(struct thread_data *td, struct fio_file *f)
 260 {
 261         struct stat st;
 262
 263         if (stat(f->file_name, &st) == -1) {
 264                 td_verror(td, errno, "fstat");
 265                 return 1;
 266         }
 267
 268         f->real_file_size = st.st_size;
 269         return 0;
 270 }
 271
 272 static int bdev_size(struct thread_data *td, struct fio_file *f)
 273 {
 274         unsigned long long bytes = 0;
 275         int r;
 276
 277         if (td->io_ops->open_file(td, f)) {
 278                 log_err("fio: failed opening blockdev %s for size check\n",
 279                         f->file_name);
 280                 return 1;
 281         }
 282
 283         r = blockdev_size(f, &bytes);
 284         if (r) {
 285                 td_verror(td, r, "blockdev_size");
 286                 goto err;
 287         }
 288
 289         if (!bytes) {
 290                 log_err("%s: zero sized block device?\n", f->file_name);
 291                 goto err;
 292         }
 293
 294         f->real_file_size = bytes;
 295         td->io_ops->close_file(td, f);
 296         return 0;
 297 err:
 298         td->io_ops->close_file(td, f);
 299         return 1;
 300 }
 301
 302 static int char_size(struct thread_data *td, struct fio_file *f)
 303 {
 304 #ifdef FIO_HAVE_CHARDEV_SIZE
 305         unsigned long long bytes = 0;
 306         int r;
 307
 308         if (td->io_ops->open_file(td, f)) {
 309                 log_err("fio: failed opening blockdev %s for size check\n",
 310                         f->file_name);
 311                 return 1;
 312         }
 313
 314         r = chardev_size(f, &bytes);
 315         if (r) {
 316                 td_verror(td, r, "chardev_size");
 317                 goto err;
 318         }
 319
 320         if (!bytes) {
 321                 log_err("%s: zero sized char device?\n", f->file_name);
 322                 goto err;
 323         }
 324
 325         f->real_file_size = bytes;
 326         td->io_ops->close_file(td, f);
 327         return 0;
 328 err:
 329         td->io_ops->close_file(td, f);
 330         return 1;
 331 #else
 332         f->real_file_size = -1ULL;
 333         return 0;
 334 #endif
 335 }
 336
 337 static int get_file_size(struct thread_data *td, struct fio_file *f)
 338 {
 339         int ret = 0;
 340
 341         if (fio_file_size_known(f))
 342                 return 0;
 343
 344         if (f->filetype == FIO_TYPE_FILE)
 345                 ret = file_size(td, f);
 346         else if (f->filetype == FIO_TYPE_BD)
 347                 ret = bdev_size(td, f);
 348         else if (f->filetype == FIO_TYPE_CHAR)
 349                 ret = char_size(td, f);
 350         else
 351                 f->real_file_size = -1;
 352
 353         if (ret)
 354                 return ret;
 355
 356         if (f->file_offset > f->real_file_size) {
 357                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 358                                         f->file_offset, f->real_file_size);
 359                 return 1;
 360         }
 361
 362         fio_file_set_size_known(f);
 363         return 0;
 364 }
 365
 366 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 367                                    unsigned long long off,
 368                                    unsigned long long len)
 369 {
 370         int ret = 0;
 371
 372         if (len == -1ULL)
 373                 len = f->io_size;
 374         if (off == -1ULL)
 375                 off = f->file_offset;
 376
 377         if (len == -1ULL || off == -1ULL)
 378                 return 0;
 379
 380         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 381                                                                 len);
 382
 383         /*
 384          * FIXME: add blockdev flushing too
 385          */
 386         if (f->mmap_ptr) {
 387                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 388 #ifdef FIO_MADV_FREE
 389                 (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 390 #endif
 391         } else if (f->filetype == FIO_TYPE_FILE) {
 392                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 393         } else if (f->filetype == FIO_TYPE_BD) {
 394                 ret = blockdev_invalidate_cache(f);
 395                 if (ret < 0 && errno == EACCES && geteuid()) {
 396                         if (!root_warn) {
 397                                 log_err("fio: only root may flush block "
 398                                         "devices. Cache flush bypassed!\n");
 399                                 root_warn = 1;
 400                         }
 401                         ret = 0;
 402                 }
 403         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 404                 ret = 0;
 405
 406         if (ret < 0) {
 407                 td_verror(td, errno, "invalidate_cache");
 408                 return 1;
 409         } else if (ret > 0) {
 410                 td_verror(td, ret, "invalidate_cache");
 411                 return 1;
 412         }
 413
 414         return ret;
 415
 416 }
 417
 418 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 419 {
 420         if (!fio_file_open(f))
 421                 return 0;
 422
 423         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 424 }
 425
 426 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 427 {
 428         int ret = 0;
 429
 430         dprint(FD_FILE, "fd close %s\n", f->file_name);
 431
 432         remove_file_hash(f);
 433
 434         if (close(f->fd) < 0)
 435                 ret = errno;
 436
 437         f->fd = -1;
 438
 439         if (f->shadow_fd != -1) {
 440                 close(f->shadow_fd);
 441                 f->shadow_fd = -1;
 442         }
 443
 444         return ret;
 445 }
 446
 447 int file_lookup_open(struct fio_file *f, int flags)
 448 {
 449         struct fio_file *__f;
 450         int from_hash;
 451
 452         __f = lookup_file_hash(f->file_name);
 453         if (__f) {
 454                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 455                 /*
 456                  * racy, need the __f->lock locked
 457                  */
 458                 f->lock = __f->lock;
 459                 from_hash = 1;
 460         } else {
 461                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 462                 from_hash = 0;
 463         }
 464
 465         f->fd = open(f->file_name, flags, 0600);
 466         return from_hash;
 467 }
 468
 469 static int file_close_shadow_fds(struct thread_data *td)
 470 {
 471         struct fio_file *f;
 472         int num_closed = 0;
 473         unsigned int i;
 474
 475         for_each_file(td, f, i) {
 476                 if (f->shadow_fd == -1)
 477                         continue;
 478
 479                 close(f->shadow_fd);
 480                 f->shadow_fd = -1;
 481                 num_closed++;
 482         }
 483
 484         return num_closed;
 485 }
 486
 487 int generic_open_file(struct thread_data *td, struct fio_file *f)
 488 {
 489         int is_std = 0;
 490         int flags = 0;
 491         int from_hash = 0;
 492
 493         dprint(FD_FILE, "fd open %s\n", f->file_name);
 494
 495         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 496                 log_err("fio: trim only applies to block device\n");
 497                 return 1;
 498         }
 499
 500         if (!strcmp(f->file_name, "-")) {
 501                 if (td_rw(td)) {
 502                         log_err("fio: can't read/write to stdin/out\n");
 503                         return 1;
 504                 }
 505                 is_std = 1;
 506
 507                 /*
 508                  * move output logging to stderr, if we are writing to stdout
 509                  */
 510                 if (td_write(td))
 511                         f_out = stderr;
 512         }
 513
 514         if (td_trim(td))
 515                 goto skip_flags;
 516         if (td->o.odirect)
 517                 flags |= OS_O_DIRECT;
 518         if (td->o.sync_io)
 519                 flags |= O_SYNC;
 520         if (td->o.create_on_open)
 521                 flags |= O_CREAT;
 522 skip_flags:
 523         if (f->filetype != FIO_TYPE_FILE)
 524                 flags |= FIO_O_NOATIME;
 525
 526 open_again:
 527         if (td_write(td)) {
 528                 if (!read_only)
 529                         flags |= O_RDWR;
 530
 531                 if (f->filetype == FIO_TYPE_FILE)
 532                         flags |= O_CREAT;
 533
 534                 if (is_std)
 535                         f->fd = dup(STDOUT_FILENO);
 536                 else
 537                         from_hash = file_lookup_open(f, flags);
 538         } else if (td_read(td)) {
 539                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 540                         flags |= O_RDWR;
 541                 else
 542                         flags |= O_RDONLY;
 543
 544                 if (is_std)
 545                         f->fd = dup(STDIN_FILENO);
 546                 else
 547                         from_hash = file_lookup_open(f, flags);
 548         } else { //td trim
 549                 flags |= O_RDWR;
 550                 from_hash = file_lookup_open(f, flags);
 551         }
 552
 553         if (f->fd == -1) {
 554                 char buf[FIO_VERROR_SIZE];
 555                 int __e = errno;
 556
 557                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 558                         flags &= ~FIO_O_NOATIME;
 559                         goto open_again;
 560                 }
 561                 if (__e == EMFILE && file_close_shadow_fds(td))
 562                         goto open_again;
 563
 564                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 565
 566                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 567                         log_err("fio: looks like your file system does not " \
 568                                 "support direct=1/buffered=0\n");
 569                 }
 570
 571                 td_verror(td, __e, buf);
 572         }
 573
 574         if (!from_hash && f->fd != -1) {
 575                 if (add_file_hash(f)) {
 576                         int fio_unused ret;
 577
 578                         /*
 579                          * Stash away descriptor for later close. This is to
 580                          * work-around a "feature" on Linux, where a close of
 581                          * an fd that has been opened for write will trigger
 582                          * udev to call blkid to check partitions, fs id, etc.
 583                          * That polutes the device cache, which can slow down
 584                          * unbuffered accesses.
 585                          */
 586                         if (f->shadow_fd == -1)
 587                                 f->shadow_fd = f->fd;
 588                         else {
 589                                 /*
 590                                  * OK to ignore, we haven't done anything
 591                                  * with it
 592                                  */
 593                                 ret = generic_close_file(td, f);
 594                         }
 595                         goto open_again;
 596                 }
 597         }
 598
 599         return 0;
 600 }
 601
 602 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 603 {
 604         return get_file_size(td, f);
 605 }
 606
 607 /*
 608  * open/close all files, so that ->real_file_size gets set
 609  */
 610 static int get_file_sizes(struct thread_data *td)
 611 {
 612         struct fio_file *f;
 613         unsigned int i;
 614         int err = 0;
 615
 616         for_each_file(td, f, i) {
 617                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 618                                                                 f->file_name);
 619
 620                 if (td_io_get_file_size(td, f)) {
 621                         if (td->error != ENOENT) {
 622                                 log_err("%s\n", td->verror);
 623                                 err = 1;
 624                         }
 625                         clear_error(td);
 626                 }
 627
 628                 if (f->real_file_size == -1ULL && td->o.size)
 629                         f->real_file_size = td->o.size / td->o.nr_files;
 630         }
 631
 632         return err;
 633 }
 634
 635 struct fio_mount {
 636         struct flist_head list;
 637         const char *base;
 638         char __base[256];
 639         unsigned int key;
 640 };
 641
 642 /*
 643  * Get free number of bytes for each file on each unique mount.
 644  */
 645 static unsigned long long get_fs_free_counts(struct thread_data *td)
 646 {
 647         struct flist_head *n, *tmp;
 648         unsigned long long ret = 0;
 649         struct fio_mount *fm;
 650         FLIST_HEAD(list);
 651         struct fio_file *f;
 652         unsigned int i;
 653
 654         for_each_file(td, f, i) {
 655                 struct stat sb;
 656                 char buf[256];
 657
 658                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 659                         if (f->real_file_size != -1ULL)
 660                                 ret += f->real_file_size;
 661                         continue;
 662                 } else if (f->filetype != FIO_TYPE_FILE)
 663                         continue;
 664
 665                 strcpy(buf, f->file_name);
 666
 667                 if (stat(buf, &sb) < 0) {
 668                         if (errno != ENOENT)
 669                                 break;
 670                         strcpy(buf, ".");
 671                         if (stat(buf, &sb) < 0)
 672                                 break;
 673                 }
 674
 675                 fm = NULL;
 676                 flist_for_each(n, &list) {
 677                         fm = flist_entry(n, struct fio_mount, list);
 678                         if (fm->key == sb.st_dev)
 679                                 break;
 680
 681                         fm = NULL;
 682                 }
 683
 684                 if (fm)
 685                         continue;
 686
 687                 fm = malloc(sizeof(*fm));
 688                 strcpy(fm->__base, buf);
 689                 fm->base = basename(fm->__base);
 690                 fm->key = sb.st_dev;
 691                 flist_add(&fm->list, &list);
 692         }
 693
 694         flist_for_each_safe(n, tmp, &list) {
 695                 unsigned long long sz;
 696
 697                 fm = flist_entry(n, struct fio_mount, list);
 698                 flist_del(&fm->list);
 699
 700                 sz = get_fs_size(fm->base);
 701                 if (sz && sz != -1ULL)
 702                         ret += sz;
 703
 704                 free(fm);
 705         }
 706
 707         return ret;
 708 }
 709
 710 uint64_t get_start_offset(struct thread_data *td)
 711 {
 712         return td->o.start_offset +
 713                 (td->thread_number - 1) * td->o.offset_increment;
 714 }
 715
 716 /*
 717  * Open the files and setup files sizes, creating files if necessary.
 718  */
 719 int setup_files(struct thread_data *td)
 720 {
 721         unsigned long long total_size, extend_size;
 722         struct thread_options *o = &td->o;
 723         struct fio_file *f;
 724         unsigned int i;
 725         int err = 0, need_extend;
 726
 727         dprint(FD_FILE, "setup files\n");
 728
 729         if (o->read_iolog_file)
 730                 goto done;
 731
 732         /*
 733          * if ioengine defines a setup() method, it's responsible for
 734          * opening the files and setting f->real_file_size to indicate
 735          * the valid range for that file.
 736          */
 737         if (td->io_ops->setup)
 738                 err = td->io_ops->setup(td);
 739         else
 740                 err = get_file_sizes(td);
 741
 742         if (err)
 743                 return err;
 744
 745         /*
 746          * check sizes. if the files/devices do not exist and the size
 747          * isn't passed to fio, abort.
 748          */
 749         total_size = 0;
 750         for_each_file(td, f, i) {
 751                 if (f->real_file_size == -1ULL)
 752                         total_size = -1ULL;
 753                 else
 754                         total_size += f->real_file_size;
 755         }
 756
 757         if (o->fill_device)
 758                 td->fill_device_size = get_fs_free_counts(td);
 759
 760         /*
 761          * device/file sizes are zero and no size given, punt
 762          */
 763         if ((!total_size || total_size == -1ULL) && !o->size &&
 764             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 765             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 766                 log_err("%s: you need to specify size=\n", o->name);
 767                 td_verror(td, EINVAL, "total_file_size");
 768                 return 1;
 769         }
 770
 771         /*
 772          * now file sizes are known, so we can set ->io_size. if size= is
 773          * not given, ->io_size is just equal to ->real_file_size. if size
 774          * is given, ->io_size is size / nr_files.
 775          */
 776         extend_size = total_size = 0;
 777         need_extend = 0;
 778         for_each_file(td, f, i) {
 779                 f->file_offset = get_start_offset(td);
 780
 781                 if (!o->file_size_low) {
 782                         /*
 783                          * no file size range given, file size is equal to
 784                          * total size divided by number of files. if that is
 785                          * zero, set it to the real file size.
 786                          */
 787                         f->io_size = o->size / o->nr_files;
 788                         if (!f->io_size)
 789                                 f->io_size = f->real_file_size - f->file_offset;
 790                 } else if (f->real_file_size < o->file_size_low ||
 791                            f->real_file_size > o->file_size_high) {
 792                         if (f->file_offset > o->file_size_low)
 793                                 goto err_offset;
 794                         /*
 795                          * file size given. if it's fixed, use that. if it's a
 796                          * range, generate a random size in-between.
 797                          */
 798                         if (o->file_size_low == o->file_size_high)
 799                                 f->io_size = o->file_size_low - f->file_offset;
 800                         else {
 801                                 f->io_size = get_rand_file_size(td)
 802                                                 - f->file_offset;
 803                         }
 804                 } else
 805                         f->io_size = f->real_file_size - f->file_offset;
 806
 807                 if (f->io_size == -1ULL)
 808                         total_size = -1ULL;
 809                 else {
 810                         if (o->size_percent)
 811                                 f->io_size = (f->io_size * o->size_percent) / 100;
 812                         total_size += f->io_size;
 813                 }
 814
 815                 if (f->filetype == FIO_TYPE_FILE &&
 816                     (f->io_size + f->file_offset) > f->real_file_size &&
 817                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 818                         if (!o->create_on_open) {
 819                                 need_extend++;
 820                                 extend_size += (f->io_size + f->file_offset);
 821                         } else
 822                                 f->real_file_size = f->io_size + f->file_offset;
 823                         fio_file_set_extend(f);
 824                 }
 825         }
 826
 827         if (!o->size || o->size > total_size)
 828                 o->size = total_size;
 829
 830         /*
 831          * See if we need to extend some files
 832          */
 833         if (need_extend) {
 834                 temp_stall_ts = 1;
 835                 if (output_format == FIO_OUTPUT_NORMAL)
 836                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 837                                  " %lluMB)\n", o->name, need_extend,
 838                                         extend_size >> 20);
 839
 840                 for_each_file(td, f, i) {
 841                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 842
 843                         if (!fio_file_extend(f))
 844                                 continue;
 845
 846                         assert(f->filetype == FIO_TYPE_FILE);
 847                         fio_file_clear_extend(f);
 848                         if (!o->fill_device) {
 849                                 old_len = f->real_file_size;
 850                                 extend_len = f->io_size + f->file_offset -
 851                                                 old_len;
 852                         }
 853                         f->real_file_size = (f->io_size + f->file_offset);
 854                         err = extend_file(td, f);
 855                         if (err)
 856                                 break;
 857
 858                         err = __file_invalidate_cache(td, f, old_len,
 859                                                                 extend_len);
 860                         close(f->fd);
 861                         f->fd = -1;
 862                         if (err)
 863                                 break;
 864                 }
 865                 temp_stall_ts = 0;
 866         }
 867
 868         if (err)
 869                 return err;
 870
 871         if (!o->zone_size)
 872                 o->zone_size = o->size;
 873
 874         /*
 875          * iolog already set the total io size, if we read back
 876          * stored entries.
 877          */
 878         if (!o->read_iolog_file)
 879                 td->total_io_size = o->size * o->loops;
 880
 881 done:
 882         if (o->create_only)
 883                 td->done = 1;
 884
 885         return 0;
 886 err_offset:
 887         log_err("%s: you need to specify valid offset=\n", o->name);
 888         return 1;
 889 }
 890
 891 int pre_read_files(struct thread_data *td)
 892 {
 893         struct fio_file *f;
 894         unsigned int i;
 895
 896         dprint(FD_FILE, "pre_read files\n");
 897
 898         for_each_file(td, f, i) {
 899                 pre_read_file(td, f);
 900         }
 901
 902         return 1;
 903 }
 904
 905 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 906 {
 907         unsigned int range_size, seed;
 908         unsigned long nranges;
 909         uint64_t file_size;
 910
 911         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 912         file_size = min(f->real_file_size, f->io_size);
 913
 914         nranges = (file_size + range_size - 1) / range_size;
 915
 916         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 917         if (!td->o.rand_repeatable)
 918                 seed = td->rand_seeds[4];
 919
 920         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 921                 zipf_init(&f->zipf, nranges, td->o.zipf_theta, seed);
 922         else
 923                 pareto_init(&f->zipf, nranges, td->o.pareto_h, seed);
 924
 925         return 1;
 926 }
 927
 928 static int init_rand_distribution(struct thread_data *td)
 929 {
 930         struct fio_file *f;
 931         unsigned int i;
 932         int state;
 933
 934         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 935                 return 0;
 936
 937         state = td->runstate;
 938         td_set_runstate(td, TD_SETTING_UP);
 939         for_each_file(td, f, i)
 940                 __init_rand_distribution(td, f);
 941         td_set_runstate(td, state);
 942
 943         return 1;
 944 }
 945
 946 int init_random_map(struct thread_data *td)
 947 {
 948         unsigned long long blocks;
 949         struct fio_file *f;
 950         unsigned int i;
 951
 952         if (init_rand_distribution(td))
 953                 return 0;
 954         if (!td_random(td))
 955                 return 0;
 956
 957         for_each_file(td, f, i) {
 958                 uint64_t file_size = min(f->real_file_size, f->io_size);
 959
 960                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
 961
 962                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
 963                         unsigned long seed;
 964
 965                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
 966
 967                         if (!lfsr_init(&f->lfsr, blocks, seed, seed & 0xF))
 968                                 continue;
 969                 } else if (!td->o.norandommap) {
 970                         f->io_axmap = axmap_new(blocks);
 971                         if (f->io_axmap)
 972                                 continue;
 973                 } else if (td->o.norandommap)
 974                         continue;
 975
 976                 if (!td->o.softrandommap) {
 977                         log_err("fio: failed allocating random map. If running"
 978                                 " a large number of jobs, try the 'norandommap'"
 979                                 " option or set 'softrandommap'. Or give"
 980                                 " a larger --alloc-size to fio.\n");
 981                         return 1;
 982                 }
 983
 984                 log_info("fio: file %s failed allocating random map. Running "
 985                          "job without.\n", f->file_name);
 986         }
 987
 988         return 0;
 989 }
 990
 991 void close_files(struct thread_data *td)
 992 {
 993         struct fio_file *f;
 994         unsigned int i;
 995
 996         for_each_file(td, f, i) {
 997                 if (fio_file_open(f))
 998                         td_io_close_file(td, f);
 999         }
1000 }
1001
1002 void close_and_free_files(struct thread_data *td)
1003 {
1004         struct fio_file *f;
1005         unsigned int i;
1006
1007         dprint(FD_FILE, "close files\n");
1008
1009         for_each_file(td, f, i) {
1010                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1011                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1012                         unlink(f->file_name);
1013                 }
1014
1015                 if (fio_file_open(f))
1016                         td_io_close_file(td, f);
1017
1018                 remove_file_hash(f);
1019
1020                 sfree(f->file_name);
1021                 f->file_name = NULL;
1022                 axmap_free(f->io_axmap);
1023                 f->io_axmap = NULL;
1024                 sfree(f);
1025         }
1026
1027         td->o.filename = NULL;
1028         free(td->files);
1029         free(td->file_locks);
1030         td->files_index = 0;
1031         td->files = NULL;
1032         td->file_locks = NULL;
1033         td->o.nr_files = 0;
1034 }
1035
1036 static void get_file_type(struct fio_file *f)
1037 {
1038         struct stat sb;
1039
1040         if (!strcmp(f->file_name, "-"))
1041                 f->filetype = FIO_TYPE_PIPE;
1042         else
1043                 f->filetype = FIO_TYPE_FILE;
1044
1045         /* \\.\ is the device namespace in Windows, where every file is
1046          * a block device */
1047         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1048                 f->filetype = FIO_TYPE_BD;
1049
1050         if (!stat(f->file_name, &sb)) {
1051                 if (S_ISBLK(sb.st_mode))
1052                         f->filetype = FIO_TYPE_BD;
1053                 else if (S_ISCHR(sb.st_mode))
1054                         f->filetype = FIO_TYPE_CHAR;
1055                 else if (S_ISFIFO(sb.st_mode))
1056                         f->filetype = FIO_TYPE_PIPE;
1057         }
1058 }
1059
1060 int add_file(struct thread_data *td, const char *fname)
1061 {
1062         int cur_files = td->files_index;
1063         char file_name[PATH_MAX];
1064         struct fio_file *f;
1065         int len = 0;
1066
1067         dprint(FD_FILE, "add file %s\n", fname);
1068
1069         f = smalloc(sizeof(*f));
1070         if (!f) {
1071                 log_err("fio: smalloc OOM\n");
1072                 assert(0);
1073         }
1074
1075         f->fd = -1;
1076         f->shadow_fd = -1;
1077         fio_file_reset(td, f);
1078
1079         if (td->files_size <= td->files_index) {
1080                 unsigned int new_size = td->o.nr_files + 1;
1081
1082                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1083
1084                 td->files = realloc(td->files, new_size * sizeof(f));
1085                 if (td->files == NULL) {
1086                         log_err("fio: realloc OOM\n");
1087                         assert(0);
1088                 }
1089                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1090                         td->file_locks = realloc(td->file_locks, new_size);
1091                         if (!td->file_locks) {
1092                                 log_err("fio: realloc OOM\n");
1093                                 assert(0);
1094                         }
1095                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1096                 }
1097                 td->files_size = new_size;
1098         }
1099         td->files[cur_files] = f;
1100         f->fileno = cur_files;
1101
1102         /*
1103          * init function, io engine may not be loaded yet
1104          */
1105         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1106                 f->real_file_size = -1ULL;
1107
1108         if (td->o.directory)
1109                 len = sprintf(file_name, "%s/", td->o.directory);
1110
1111         sprintf(file_name + len, "%s", fname);
1112         f->file_name = smalloc_strdup(file_name);
1113         if (!f->file_name) {
1114                 log_err("fio: smalloc OOM\n");
1115                 assert(0);
1116         }
1117
1118         get_file_type(f);
1119
1120         switch (td->o.file_lock_mode) {
1121         case FILE_LOCK_NONE:
1122                 break;
1123         case FILE_LOCK_READWRITE:
1124                 f->rwlock = fio_rwlock_init();
1125                 break;
1126         case FILE_LOCK_EXCLUSIVE:
1127                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1128                 break;
1129         default:
1130                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1131                 assert(0);
1132         }
1133
1134         td->files_index++;
1135         if (f->filetype == FIO_TYPE_FILE)
1136                 td->nr_normal_files++;
1137
1138         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1139                                                         cur_files);
1140
1141         return cur_files;
1142 }
1143
1144 int add_file_exclusive(struct thread_data *td, const char *fname)
1145 {
1146         struct fio_file *f;
1147         unsigned int i;
1148
1149         for_each_file(td, f, i) {
1150                 if (!strcmp(f->file_name, fname))
1151                         return i;
1152         }
1153
1154         return add_file(td, fname);
1155 }
1156
1157 void get_file(struct fio_file *f)
1158 {
1159         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1160         assert(fio_file_open(f));
1161         f->references++;
1162 }
1163
1164 int put_file(struct thread_data *td, struct fio_file *f)
1165 {
1166         int f_ret = 0, ret = 0;
1167
1168         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1169
1170         if (!fio_file_open(f)) {
1171                 assert(f->fd == -1);
1172                 return 0;
1173         }
1174
1175         assert(f->references);
1176         if (--f->references)
1177                 return 0;
1178
1179         if (should_fsync(td) && td->o.fsync_on_close)
1180                 f_ret = fsync(f->fd);
1181
1182         if (td->io_ops->close_file)
1183                 ret = td->io_ops->close_file(td, f);
1184
1185         if (!ret)
1186                 ret = f_ret;
1187
1188         td->nr_open_files--;
1189         fio_file_clear_open(f);
1190         assert(f->fd == -1);
1191         return ret;
1192 }
1193
1194 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1195 {
1196         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1197                 return;
1198
1199         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1200                 if (ddir == DDIR_READ)
1201                         fio_rwlock_read(f->rwlock);
1202                 else
1203                         fio_rwlock_write(f->rwlock);
1204         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1205                 fio_mutex_down(f->lock);
1206
1207         td->file_locks[f->fileno] = td->o.file_lock_mode;
1208 }
1209
1210 void unlock_file(struct thread_data *td, struct fio_file *f)
1211 {
1212         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1213                 return;
1214
1215         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1216                 fio_rwlock_unlock(f->rwlock);
1217         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1218                 fio_mutex_up(f->lock);
1219
1220         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1221 }
1222
1223 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1224 {
1225         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1226                 unlock_file(td, f);
1227 }
1228
1229 static int recurse_dir(struct thread_data *td, const char *dirname)
1230 {
1231         struct dirent *dir;
1232         int ret = 0;
1233         DIR *D;
1234
1235         D = opendir(dirname);
1236         if (!D) {
1237                 char buf[FIO_VERROR_SIZE];
1238
1239                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1240                 td_verror(td, errno, buf);
1241                 return 1;
1242         }
1243
1244         while ((dir = readdir(D)) != NULL) {
1245                 char full_path[PATH_MAX];
1246                 struct stat sb;
1247
1248                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1249                         continue;
1250
1251                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1252
1253                 if (lstat(full_path, &sb) == -1) {
1254                         if (errno != ENOENT) {
1255                                 td_verror(td, errno, "stat");
1256                                 return 1;
1257                         }
1258                 }
1259
1260                 if (S_ISREG(sb.st_mode)) {
1261                         add_file(td, full_path);
1262                         td->o.nr_files++;
1263                         continue;
1264                 }
1265                 if (!S_ISDIR(sb.st_mode))
1266                         continue;
1267
1268                 ret = recurse_dir(td, full_path);
1269                 if (ret)
1270                         break;
1271         }
1272
1273         closedir(D);
1274         return ret;
1275 }
1276
1277 int add_dir_files(struct thread_data *td, const char *path)
1278 {
1279         int ret = recurse_dir(td, path);
1280
1281         if (!ret)
1282                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1283
1284         return ret;
1285 }
1286
1287 void dup_files(struct thread_data *td, struct thread_data *org)
1288 {
1289         struct fio_file *f;
1290         unsigned int i;
1291
1292         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1293
1294         if (!org->files)
1295                 return;
1296
1297         td->files = malloc(org->files_index * sizeof(f));
1298
1299         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1300                 td->file_locks = malloc(org->files_index);
1301
1302         for_each_file(org, f, i) {
1303                 struct fio_file *__f;
1304
1305                 __f = smalloc(sizeof(*__f));
1306                 if (!__f) {
1307                         log_err("fio: smalloc OOM\n");
1308                         assert(0);
1309                 }
1310                 __f->fd = -1;
1311                 fio_file_reset(td, __f);
1312
1313                 if (f->file_name) {
1314                         __f->file_name = smalloc_strdup(f->file_name);
1315                         if (!__f->file_name) {
1316                                 log_err("fio: smalloc OOM\n");
1317                                 assert(0);
1318                         }
1319
1320                         __f->filetype = f->filetype;
1321                 }
1322
1323                 td->files[i] = __f;
1324         }
1325 }
1326
1327 /*
1328  * Returns the index that matches the filename, or -1 if not there
1329  */
1330 int get_fileno(struct thread_data *td, const char *fname)
1331 {
1332         struct fio_file *f;
1333         unsigned int i;
1334
1335         for_each_file(td, f, i)
1336                 if (!strcmp(f->file_name, fname))
1337                         return i;
1338
1339         return -1;
1340 }
1341
1342 /*
1343  * For log usage, where we add/open/close files automatically
1344  */
1345 void free_release_files(struct thread_data *td)
1346 {
1347         close_files(td);
1348         td->files_index = 0;
1349         td->nr_normal_files = 0;
1350 }
1351
1352 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1353 {
1354         f->last_pos = f->file_offset;
1355         f->last_start = -1ULL;
1356         if (f->io_axmap)
1357                 axmap_reset(f->io_axmap);
1358         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1359                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1360 }