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