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 "options.h"
  15 #include "os/os.h"
  16 #include "hash.h"
  17 #include "lib/axmap.h"
  18
  19 #ifdef CONFIG_LINUX_FALLOCATE
  20 #include <linux/falloc.h>
  21 #endif
  22
  23 static int root_warn;
  24
  25 static FLIST_HEAD(filename_list);
  26
  27 /*
  28  * List entry for filename_list
  29  */
  30 struct file_name {
  31         struct flist_head list;
  32         char *filename;
  33 };
  34
  35 static inline void clear_error(struct thread_data *td)
  36 {
  37         td->error = 0;
  38         td->verror[0] = '\0';
  39 }
  40
  41 /*
  42  * Leaves f->fd open on success, caller must close
  43  */
  44 static int extend_file(struct thread_data *td, struct fio_file *f)
  45 {
  46         int r, new_layout = 0, unlink_file = 0, flags;
  47         unsigned long long left;
  48         unsigned int bs;
  49         char *b = NULL;
  50
  51         if (read_only) {
  52                 log_err("fio: refusing extend of file due to read-only\n");
  53                 return 0;
  54         }
  55
  56         /*
  57          * check if we need to lay the file out complete again. fio
  58          * does that for operations involving reads, or for writes
  59          * where overwrite is set
  60          */
  61         if (td_read(td) ||
  62            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
  63             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
  64                 new_layout = 1;
  65         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
  66                 unlink_file = 1;
  67
  68         if (unlink_file || new_layout) {
  69                 int ret;
  70
  71                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  72
  73                 ret = td_io_unlink_file(td, f);
  74                 if (ret != 0 && ret != ENOENT) {
  75                         td_verror(td, errno, "unlink");
  76                         return 1;
  77                 }
  78         }
  79
  80         flags = O_WRONLY;
  81         if (td->o.allow_create)
  82                 flags |= O_CREAT;
  83         if (new_layout)
  84                 flags |= O_TRUNC;
  85
  86 #ifdef WIN32
  87         flags |= _O_BINARY;
  88 #endif
  89
  90         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  91         f->fd = open(f->file_name, flags, 0644);
  92         if (f->fd < 0) {
  93                 int err = errno;
  94
  95                 if (err == ENOENT && !td->o.allow_create)
  96                         log_err("fio: file creation disallowed by "
  97                                         "allow_file_create=0\n");
  98                 else
  99                         td_verror(td, err, "open");
 100                 return 1;
 101         }
 102
 103 #ifdef CONFIG_POSIX_FALLOCATE
 104         if (!td->o.fill_device) {
 105                 switch (td->o.fallocate_mode) {
 106                 case FIO_FALLOCATE_NONE:
 107                         break;
 108                 case FIO_FALLOCATE_POSIX:
 109                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
 110                                  f->file_name,
 111                                  (unsigned long long) f->real_file_size);
 112
 113                         r = posix_fallocate(f->fd, 0, f->real_file_size);
 114                         if (r > 0) {
 115                                 log_err("fio: posix_fallocate fails: %s\n",
 116                                                 strerror(r));
 117                         }
 118                         break;
 119 #ifdef CONFIG_LINUX_FALLOCATE
 120                 case FIO_FALLOCATE_KEEP_SIZE:
 121                         dprint(FD_FILE,
 122                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
 123                                 "file %s size %llu\n", f->file_name,
 124                                 (unsigned long long) f->real_file_size);
 125
 126                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
 127                                         f->real_file_size);
 128                         if (r != 0)
 129                                 td_verror(td, errno, "fallocate");
 130
 131                         break;
 132 #endif /* CONFIG_LINUX_FALLOCATE */
 133                 default:
 134                         log_err("fio: unknown fallocate mode: %d\n",
 135                                 td->o.fallocate_mode);
 136                         assert(0);
 137                 }
 138         }
 139 #endif /* CONFIG_POSIX_FALLOCATE */
 140
 141         /*
 142          * If our jobs don't require regular files initially, we're done.
 143          */
 144         if (!new_layout)
 145                 goto done;
 146
 147         /*
 148          * The size will be -1ULL when fill_device is used, so don't truncate
 149          * or fallocate this file, just write it
 150          */
 151         if (!td->o.fill_device) {
 152                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 153                                         (unsigned long long) f->real_file_size);
 154                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 155                         if (errno != EFBIG) {
 156                                 td_verror(td, errno, "ftruncate");
 157                                 goto err;
 158                         }
 159                 }
 160         }
 161
 162         left = f->real_file_size;
 163         bs = td->o.max_bs[DDIR_WRITE];
 164         if (bs > left)
 165                 bs = left;
 166
 167         b = malloc(bs);
 168         if (!b) {
 169                 td_verror(td, errno, "malloc");
 170                 goto err;
 171         }
 172
 173         while (left && !td->terminate) {
 174                 if (bs > left)
 175                         bs = left;
 176
 177                 fill_io_buffer(td, b, bs, bs);
 178
 179                 r = write(f->fd, b, bs);
 180
 181                 if (r > 0) {
 182                         left -= r;
 183                         continue;
 184                 } else {
 185                         if (r < 0) {
 186                                 int __e = errno;
 187
 188                                 if (__e == ENOSPC) {
 189                                         if (td->o.fill_device)
 190                                                 break;
 191                                         log_info("fio: ENOSPC on laying out "
 192                                                  "file, stopping\n");
 193                                         break;
 194                                 }
 195                                 td_verror(td, errno, "write");
 196                         } else
 197                                 td_verror(td, EIO, "write");
 198
 199                         break;
 200                 }
 201         }
 202
 203         if (td->terminate) {
 204                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 205                 td_io_unlink_file(td, f);
 206         } else if (td->o.create_fsync) {
 207                 if (fsync(f->fd) < 0) {
 208                         td_verror(td, errno, "fsync");
 209                         goto err;
 210                 }
 211         }
 212         if (td->o.fill_device && !td_write(td)) {
 213                 fio_file_clear_size_known(f);
 214                 if (td_io_get_file_size(td, f))
 215                         goto err;
 216                 if (f->io_size > f->real_file_size)
 217                         f->io_size = f->real_file_size;
 218         }
 219
 220         free(b);
 221 done:
 222         return 0;
 223 err:
 224         close(f->fd);
 225         f->fd = -1;
 226         if (b)
 227                 free(b);
 228         return 1;
 229 }
 230
 231 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 232 {
 233         int ret = 0, r, did_open = 0, old_runstate;
 234         unsigned long long left;
 235         unsigned int bs;
 236         char *b;
 237
 238         if (td_ioengine_flagged(td, FIO_PIPEIO))
 239                 return 0;
 240
 241         if (!fio_file_open(f)) {
 242                 if (td->io_ops->open_file(td, f)) {
 243                         log_err("fio: cannot pre-read, failed to open file\n");
 244                         return 1;
 245                 }
 246                 did_open = 1;
 247         }
 248
 249         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 250
 251         left = f->io_size;
 252         bs = td->o.max_bs[DDIR_READ];
 253         if (bs > left)
 254                 bs = left;
 255
 256         b = malloc(bs);
 257         if (!b) {
 258                 td_verror(td, errno, "malloc");
 259                 ret = 1;
 260                 goto error;
 261         }
 262         memset(b, 0, bs);
 263
 264         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 265                 td_verror(td, errno, "lseek");
 266                 log_err("fio: failed to lseek pre-read file\n");
 267                 ret = 1;
 268                 goto error;
 269         }
 270
 271         while (left && !td->terminate) {
 272                 if (bs > left)
 273                         bs = left;
 274
 275                 r = read(f->fd, b, bs);
 276
 277                 if (r == (int) bs) {
 278                         left -= bs;
 279                         continue;
 280                 } else {
 281                         td_verror(td, EIO, "pre_read");
 282                         break;
 283                 }
 284         }
 285
 286 error:
 287         td_restore_runstate(td, old_runstate);
 288
 289         if (did_open)
 290                 td->io_ops->close_file(td, f);
 291
 292         free(b);
 293         return ret;
 294 }
 295
 296 unsigned long long get_rand_file_size(struct thread_data *td)
 297 {
 298         unsigned long long ret, sized;
 299         uint64_t frand_max;
 300         unsigned long r;
 301
 302         frand_max = rand_max(&td->file_size_state);
 303         r = __rand(&td->file_size_state);
 304         sized = td->o.file_size_high - td->o.file_size_low;
 305         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 306         ret += td->o.file_size_low;
 307         ret -= (ret % td->o.rw_min_bs);
 308         return ret;
 309 }
 310
 311 static int file_size(struct thread_data *td, struct fio_file *f)
 312 {
 313         struct stat st;
 314
 315         if (stat(f->file_name, &st) == -1) {
 316                 td_verror(td, errno, "fstat");
 317                 return 1;
 318         }
 319
 320         f->real_file_size = st.st_size;
 321         return 0;
 322 }
 323
 324 static int bdev_size(struct thread_data *td, struct fio_file *f)
 325 {
 326         unsigned long long bytes = 0;
 327         int r;
 328
 329         if (td->io_ops->open_file(td, f)) {
 330                 log_err("fio: failed opening blockdev %s for size check\n",
 331                         f->file_name);
 332                 return 1;
 333         }
 334
 335         r = blockdev_size(f, &bytes);
 336         if (r) {
 337                 td_verror(td, r, "blockdev_size");
 338                 goto err;
 339         }
 340
 341         if (!bytes) {
 342                 log_err("%s: zero sized block device?\n", f->file_name);
 343                 goto err;
 344         }
 345
 346         f->real_file_size = bytes;
 347         td->io_ops->close_file(td, f);
 348         return 0;
 349 err:
 350         td->io_ops->close_file(td, f);
 351         return 1;
 352 }
 353
 354 static int char_size(struct thread_data *td, struct fio_file *f)
 355 {
 356 #ifdef FIO_HAVE_CHARDEV_SIZE
 357         unsigned long long bytes = 0;
 358         int r;
 359
 360         if (td->io_ops->open_file(td, f)) {
 361                 log_err("fio: failed opening chardev %s for size check\n",
 362                         f->file_name);
 363                 return 1;
 364         }
 365
 366         r = chardev_size(f, &bytes);
 367         if (r) {
 368                 td_verror(td, r, "chardev_size");
 369                 goto err;
 370         }
 371
 372         if (!bytes) {
 373                 log_err("%s: zero sized char device?\n", f->file_name);
 374                 goto err;
 375         }
 376
 377         f->real_file_size = bytes;
 378         td->io_ops->close_file(td, f);
 379         return 0;
 380 err:
 381         td->io_ops->close_file(td, f);
 382         return 1;
 383 #else
 384         f->real_file_size = -1ULL;
 385         return 0;
 386 #endif
 387 }
 388
 389 static int get_file_size(struct thread_data *td, struct fio_file *f)
 390 {
 391         int ret = 0;
 392
 393         if (fio_file_size_known(f))
 394                 return 0;
 395
 396         if (f->filetype == FIO_TYPE_FILE)
 397                 ret = file_size(td, f);
 398         else if (f->filetype == FIO_TYPE_BLOCK)
 399                 ret = bdev_size(td, f);
 400         else if (f->filetype == FIO_TYPE_CHAR)
 401                 ret = char_size(td, f);
 402         else
 403                 f->real_file_size = -1ULL;
 404
 405         /*
 406          * Leave ->real_file_size with 0 since it could be expectation
 407          * of initial setup for regular files.
 408          */
 409         if (ret)
 410                 return ret;
 411
 412         /*
 413          * If ->real_file_size is -1, a conditional for the message
 414          * "offset extends end" is always true, but it makes no sense,
 415          * so just return the same value here.
 416          */
 417         if (f->real_file_size == -1ULL) {
 418                 log_info("%s: failed to get file size of %s\n", td->o.name,
 419                                         f->file_name);
 420                 return 1;
 421         }
 422
 423         if (td->o.start_offset && f->file_offset == 0)
 424                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 425                                         f->file_name);
 426         /*
 427          * ->file_offset normally hasn't been initialized yet, so this
 428          * is basically always false.
 429          */
 430         if (f->file_offset > f->real_file_size) {
 431                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 432                                         (unsigned long long) f->file_offset,
 433                                         (unsigned long long) f->real_file_size);
 434                 return 1;
 435         }
 436
 437         fio_file_set_size_known(f);
 438         return 0;
 439 }
 440
 441 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 442                                    unsigned long long off,
 443                                    unsigned long long len)
 444 {
 445         int errval = 0, ret = 0;
 446
 447 #ifdef CONFIG_ESX
 448         return 0;
 449 #endif
 450
 451         if (len == -1ULL)
 452                 len = f->io_size;
 453         if (off == -1ULL)
 454                 off = f->file_offset;
 455
 456         if (len == -1ULL || off == -1ULL)
 457                 return 0;
 458
 459         if (td->io_ops->invalidate) {
 460                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 461                         f->file_name);
 462                 ret = td->io_ops->invalidate(td, f);
 463                 if (ret < 0)
 464                         errval = -ret;
 465         } else if (f->filetype == FIO_TYPE_FILE) {
 466                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 467                         f->file_name, off, len);
 468                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 469                 if (ret)
 470                         errval = ret;
 471         } else if (f->filetype == FIO_TYPE_BLOCK) {
 472                 int retry_count = 0;
 473
 474                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 475                 ret = blockdev_invalidate_cache(f);
 476                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 477                         /*
 478                          * Linux multipath devices reject ioctl while
 479                          * the maps are being updated. That window can
 480                          * last tens of milliseconds; we'll try up to
 481                          * a quarter of a second.
 482                          */
 483                         usleep(10000);
 484                         ret = blockdev_invalidate_cache(f);
 485                 }
 486                 if (ret < 0 && errno == EACCES && geteuid()) {
 487                         if (!root_warn) {
 488                                 log_err("fio: only root may flush block "
 489                                         "devices. Cache flush bypassed!\n");
 490                                 root_warn = 1;
 491                         }
 492                         ret = 0;
 493                 }
 494                 if (ret < 0)
 495                         errval = errno;
 496                 else if (ret) /* probably not supported */
 497                         errval = ret;
 498         } else if (f->filetype == FIO_TYPE_CHAR ||
 499                    f->filetype == FIO_TYPE_PIPE) {
 500                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 501                 ret = 0;
 502         }
 503
 504         /*
 505          * Cache flushing isn't a fatal condition, and we know it will
 506          * happen on some platforms where we don't have the proper
 507          * function to flush eg block device caches. So just warn and
 508          * continue on our way.
 509          */
 510         if (errval)
 511                 log_info("fio: cache invalidation of %s failed: %s\n",
 512                          f->file_name, strerror(errval));
 513
 514         return 0;
 515
 516 }
 517
 518 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 519 {
 520         if (!fio_file_open(f))
 521                 return 0;
 522
 523         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 524 }
 525
 526 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 527 {
 528         int ret = 0;
 529
 530         dprint(FD_FILE, "fd close %s\n", f->file_name);
 531
 532         remove_file_hash(f);
 533
 534         if (close(f->fd) < 0)
 535                 ret = errno;
 536
 537         f->fd = -1;
 538
 539         if (f->shadow_fd != -1) {
 540                 close(f->shadow_fd);
 541                 f->shadow_fd = -1;
 542         }
 543
 544         f->engine_pos = 0;
 545         return ret;
 546 }
 547
 548 int file_lookup_open(struct fio_file *f, int flags)
 549 {
 550         struct fio_file *__f;
 551         int from_hash;
 552
 553         __f = lookup_file_hash(f->file_name);
 554         if (__f) {
 555                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 556                 f->lock = __f->lock;
 557                 from_hash = 1;
 558         } else {
 559                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 560                 from_hash = 0;
 561         }
 562
 563 #ifdef WIN32
 564         flags |= _O_BINARY;
 565 #endif
 566
 567         f->fd = open(f->file_name, flags, 0600);
 568         return from_hash;
 569 }
 570
 571 static int file_close_shadow_fds(struct thread_data *td)
 572 {
 573         struct fio_file *f;
 574         int num_closed = 0;
 575         unsigned int i;
 576
 577         for_each_file(td, f, i) {
 578                 if (f->shadow_fd == -1)
 579                         continue;
 580
 581                 close(f->shadow_fd);
 582                 f->shadow_fd = -1;
 583                 num_closed++;
 584         }
 585
 586         return num_closed;
 587 }
 588
 589 int generic_open_file(struct thread_data *td, struct fio_file *f)
 590 {
 591         int is_std = 0;
 592         int flags = 0;
 593         int from_hash = 0;
 594
 595         dprint(FD_FILE, "fd open %s\n", f->file_name);
 596
 597         if (!strcmp(f->file_name, "-")) {
 598                 if (td_rw(td)) {
 599                         log_err("fio: can't read/write to stdin/out\n");
 600                         return 1;
 601                 }
 602                 is_std = 1;
 603
 604                 /*
 605                  * move output logging to stderr, if we are writing to stdout
 606                  */
 607                 if (td_write(td))
 608                         f_out = stderr;
 609         }
 610
 611         if (td_trim(td))
 612                 goto skip_flags;
 613         if (td->o.odirect)
 614                 flags |= OS_O_DIRECT;
 615         if (td->o.oatomic) {
 616                 if (!FIO_O_ATOMIC) {
 617                         td_verror(td, EINVAL, "OS does not support atomic IO");
 618                         return 1;
 619                 }
 620                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 621         }
 622         if (td->o.sync_io)
 623                 flags |= O_SYNC;
 624         if (td->o.create_on_open && td->o.allow_create)
 625                 flags |= O_CREAT;
 626 skip_flags:
 627         if (f->filetype != FIO_TYPE_FILE)
 628                 flags |= FIO_O_NOATIME;
 629
 630 open_again:
 631         if (td_write(td)) {
 632                 if (!read_only)
 633                         flags |= O_RDWR;
 634
 635                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 636                         flags |= O_CREAT;
 637
 638                 if (is_std)
 639                         f->fd = dup(STDOUT_FILENO);
 640                 else
 641                         from_hash = file_lookup_open(f, flags);
 642         } else if (td_read(td)) {
 643                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 644                         flags |= O_RDWR;
 645                 else
 646                         flags |= O_RDONLY;
 647
 648                 if (is_std)
 649                         f->fd = dup(STDIN_FILENO);
 650                 else
 651                         from_hash = file_lookup_open(f, flags);
 652         } else if (td_trim(td)) {
 653                 assert(!td_rw(td)); /* should have matched above */
 654                 flags |= O_RDWR;
 655                 from_hash = file_lookup_open(f, flags);
 656         }
 657
 658         if (f->fd == -1) {
 659                 char buf[FIO_VERROR_SIZE];
 660                 int __e = errno;
 661
 662                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 663                         flags &= ~FIO_O_NOATIME;
 664                         goto open_again;
 665                 }
 666                 if (__e == EMFILE && file_close_shadow_fds(td))
 667                         goto open_again;
 668
 669                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 670
 671                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 672                         log_err("fio: looks like your file system does not " \
 673                                 "support direct=1/buffered=0\n");
 674                 }
 675
 676                 td_verror(td, __e, buf);
 677                 return 1;
 678         }
 679
 680         if (!from_hash && f->fd != -1) {
 681                 if (add_file_hash(f)) {
 682                         int fio_unused ret;
 683
 684                         /*
 685                          * Stash away descriptor for later close. This is to
 686                          * work-around a "feature" on Linux, where a close of
 687                          * an fd that has been opened for write will trigger
 688                          * udev to call blkid to check partitions, fs id, etc.
 689                          * That pollutes the device cache, which can slow down
 690                          * unbuffered accesses.
 691                          */
 692                         if (f->shadow_fd == -1)
 693                                 f->shadow_fd = f->fd;
 694                         else {
 695                                 /*
 696                                  * OK to ignore, we haven't done anything
 697                                  * with it
 698                                  */
 699                                 ret = generic_close_file(td, f);
 700                         }
 701                         goto open_again;
 702                 }
 703         }
 704
 705         return 0;
 706 }
 707
 708 /*
 709  * This function i.e. get_file_size() is the default .get_file_size
 710  * implementation of majority of I/O engines.
 711  */
 712 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 713 {
 714         return get_file_size(td, f);
 715 }
 716
 717 /*
 718  * open/close all files, so that ->real_file_size gets set
 719  */
 720 static int get_file_sizes(struct thread_data *td)
 721 {
 722         struct fio_file *f;
 723         unsigned int i;
 724         int err = 0;
 725
 726         for_each_file(td, f, i) {
 727                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 728                                                                 f->file_name);
 729
 730                 if (td_io_get_file_size(td, f)) {
 731                         if (td->error != ENOENT) {
 732                                 log_err("%s\n", td->verror);
 733                                 err = 1;
 734                                 break;
 735                         }
 736                         clear_error(td);
 737                 }
 738
 739                 /*
 740                  * There are corner cases where we end up with -1 for
 741                  * ->real_file_size due to unsupported file type, etc.
 742                  * We then just set to size option value divided by number
 743                  * of files, similar to the way file ->io_size is set.
 744                  * stat(2) failure doesn't set ->real_file_size to -1.
 745                  */
 746                 if (f->real_file_size == -1ULL && td->o.size)
 747                         f->real_file_size = td->o.size / td->o.nr_files;
 748         }
 749
 750         return err;
 751 }
 752
 753 struct fio_mount {
 754         struct flist_head list;
 755         const char *base;
 756         char __base[256];
 757         unsigned int key;
 758 };
 759
 760 /*
 761  * Get free number of bytes for each file on each unique mount.
 762  */
 763 static unsigned long long get_fs_free_counts(struct thread_data *td)
 764 {
 765         struct flist_head *n, *tmp;
 766         unsigned long long ret = 0;
 767         struct fio_mount *fm;
 768         FLIST_HEAD(list);
 769         struct fio_file *f;
 770         unsigned int i;
 771
 772         for_each_file(td, f, i) {
 773                 struct stat sb;
 774                 char buf[256];
 775
 776                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 777                         if (f->real_file_size != -1ULL)
 778                                 ret += f->real_file_size;
 779                         continue;
 780                 } else if (f->filetype != FIO_TYPE_FILE)
 781                         continue;
 782
 783                 buf[255] = '\0';
 784                 strncpy(buf, f->file_name, 255);
 785
 786                 if (stat(buf, &sb) < 0) {
 787                         if (errno != ENOENT)
 788                                 break;
 789                         strcpy(buf, ".");
 790                         if (stat(buf, &sb) < 0)
 791                                 break;
 792                 }
 793
 794                 fm = NULL;
 795                 flist_for_each(n, &list) {
 796                         fm = flist_entry(n, struct fio_mount, list);
 797                         if (fm->key == sb.st_dev)
 798                                 break;
 799
 800                         fm = NULL;
 801                 }
 802
 803                 if (fm)
 804                         continue;
 805
 806                 fm = calloc(1, sizeof(*fm));
 807                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 808                 fm->base = basename(fm->__base);
 809                 fm->key = sb.st_dev;
 810                 flist_add(&fm->list, &list);
 811         }
 812
 813         flist_for_each_safe(n, tmp, &list) {
 814                 unsigned long long sz;
 815
 816                 fm = flist_entry(n, struct fio_mount, list);
 817                 flist_del(&fm->list);
 818
 819                 sz = get_fs_free_size(fm->base);
 820                 if (sz && sz != -1ULL)
 821                         ret += sz;
 822
 823                 free(fm);
 824         }
 825
 826         return ret;
 827 }
 828
 829 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 830 {
 831         struct thread_options *o = &td->o;
 832
 833         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 834                 return f->real_file_size;
 835
 836         return td->o.start_offset +
 837                 td->subjob_number * td->o.offset_increment;
 838 }
 839
 840 /*
 841  * Open the files and setup files sizes, creating files if necessary.
 842  */
 843 int setup_files(struct thread_data *td)
 844 {
 845         unsigned long long total_size, extend_size;
 846         struct thread_options *o = &td->o;
 847         struct fio_file *f;
 848         unsigned int i, nr_fs_extra = 0;
 849         int err = 0, need_extend;
 850         int old_state;
 851         const unsigned int bs = td_min_bs(td);
 852         uint64_t fs = 0;
 853
 854         dprint(FD_FILE, "setup files\n");
 855
 856         old_state = td_bump_runstate(td, TD_SETTING_UP);
 857
 858         if (o->read_iolog_file)
 859                 goto done;
 860
 861         /*
 862          * Find out physical size of files or devices for this thread,
 863          * before we determine I/O size and range of our targets.
 864          * If ioengine defines a setup() method, it's responsible for
 865          * opening the files and setting f->real_file_size to indicate
 866          * the valid range for that file.
 867          */
 868         if (td->io_ops->setup)
 869                 err = td->io_ops->setup(td);
 870         else
 871                 err = get_file_sizes(td);
 872
 873         if (err)
 874                 goto err_out;
 875
 876         /*
 877          * check sizes. if the files/devices do not exist and the size
 878          * isn't passed to fio, abort.
 879          */
 880         total_size = 0;
 881         for_each_file(td, f, i) {
 882                 f->fileno = i;
 883                 if (f->real_file_size == -1ULL)
 884                         total_size = -1ULL;
 885                 else
 886                         total_size += f->real_file_size;
 887         }
 888
 889         if (o->fill_device)
 890                 td->fill_device_size = get_fs_free_counts(td);
 891
 892         /*
 893          * device/file sizes are zero and no size given, punt
 894          */
 895         if ((!total_size || total_size == -1ULL) && !o->size &&
 896             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 897             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 898                 log_err("%s: you need to specify size=\n", o->name);
 899                 td_verror(td, EINVAL, "total_file_size");
 900                 goto err_out;
 901         }
 902
 903         /*
 904          * Calculate per-file size and potential extra size for the
 905          * first files, if needed (i.e. if we don't have a fixed size).
 906          */
 907         if (!o->file_size_low && o->nr_files) {
 908                 uint64_t all_fs;
 909
 910                 fs = o->size / o->nr_files;
 911                 all_fs = fs * o->nr_files;
 912
 913                 if (all_fs < o->size)
 914                         nr_fs_extra = (o->size - all_fs) / bs;
 915         }
 916
 917         /*
 918          * now file sizes are known, so we can set ->io_size. if size= is
 919          * not given, ->io_size is just equal to ->real_file_size. if size
 920          * is given, ->io_size is size / nr_files.
 921          */
 922         extend_size = total_size = 0;
 923         need_extend = 0;
 924         for_each_file(td, f, i) {
 925                 f->file_offset = get_start_offset(td, f);
 926
 927                 /*
 928                  * Update ->io_size depending on options specified.
 929                  * ->file_size_low being 0 means filesize option isn't set.
 930                  * Non zero ->file_size_low equals ->file_size_high means
 931                  * filesize option is set in a fixed size format.
 932                  * Non zero ->file_size_low not equals ->file_size_high means
 933                  * filesize option is set in a range format.
 934                  */
 935                 if (!o->file_size_low) {
 936                         /*
 937                          * no file size or range given, file size is equal to
 938                          * total size divided by number of files. If the size
 939                          * doesn't divide nicely with the min blocksize,
 940                          * make the first files bigger.
 941                          */
 942                         f->io_size = fs;
 943                         if (nr_fs_extra) {
 944                                 nr_fs_extra--;
 945                                 f->io_size += bs;
 946                         }
 947
 948                         /*
 949                          * We normally don't come here for regular files, but
 950                          * if the result is 0 for a regular file, set it to the
 951                          * real file size. This could be size of the existing
 952                          * one if it already exists, but otherwise will be set
 953                          * to 0. A new file won't be created because
 954                          * ->io_size + ->file_offset equals ->real_file_size.
 955                          */
 956                         if (!f->io_size) {
 957                                 if (f->file_offset > f->real_file_size)
 958                                         goto err_offset;
 959                                 f->io_size = f->real_file_size - f->file_offset;
 960                                 if (!f->io_size)
 961                                         log_info("fio: file %s may be ignored\n",
 962                                                 f->file_name);
 963                         }
 964                 } else if (f->real_file_size < o->file_size_low ||
 965                            f->real_file_size > o->file_size_high) {
 966                         if (f->file_offset > o->file_size_low)
 967                                 goto err_offset;
 968                         /*
 969                          * file size given. if it's fixed, use that. if it's a
 970                          * range, generate a random size in-between.
 971                          */
 972                         if (o->file_size_low == o->file_size_high)
 973                                 f->io_size = o->file_size_low - f->file_offset;
 974                         else {
 975                                 f->io_size = get_rand_file_size(td)
 976                                                 - f->file_offset;
 977                         }
 978                 } else
 979                         f->io_size = f->real_file_size - f->file_offset;
 980
 981                 if (f->io_size == -1ULL)
 982                         total_size = -1ULL;
 983                 else {
 984                         if (o->size_percent) {
 985                                 f->io_size = (f->io_size * o->size_percent) / 100;
 986                                 f->io_size -= (f->io_size % td_min_bs(td));
 987                         }
 988                         total_size += f->io_size;
 989                 }
 990
 991                 if (f->filetype == FIO_TYPE_FILE &&
 992                     (f->io_size + f->file_offset) > f->real_file_size &&
 993                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 994                         if (!o->create_on_open) {
 995                                 need_extend++;
 996                                 extend_size += (f->io_size + f->file_offset);
 997                                 fio_file_set_extend(f);
 998                         } else
 999                                 f->real_file_size = f->io_size + f->file_offset;
1000                 }
1001         }
1002
1003         if (td->o.block_error_hist) {
1004                 int len;
1005
1006                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1007                 f = td->files[0];
1008                 len = f->io_size / td->o.bs[DDIR_TRIM];
1009                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1010                         log_err("fio: cannot calculate block histogram with "
1011                                 "%d trim blocks, maximum %d\n",
1012                                 len, MAX_NR_BLOCK_INFOS);
1013                         td_verror(td, EINVAL, "block_error_hist");
1014                         goto err_out;
1015                 }
1016
1017                 td->ts.nr_block_infos = len;
1018                 for (i = 0; i < len; i++)
1019                         td->ts.block_infos[i] =
1020                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1021         } else
1022                 td->ts.nr_block_infos = 0;
1023
1024         if (!o->size || (total_size && o->size > total_size))
1025                 o->size = total_size;
1026
1027         if (o->size < td_min_bs(td)) {
1028                 log_err("fio: blocksize too large for data set\n");
1029                 goto err_out;
1030         }
1031
1032         /*
1033          * See if we need to extend some files, typically needed when our
1034          * target regular files don't exist yet, but our jobs require them
1035          * initially due to read I/Os.
1036          */
1037         if (need_extend) {
1038                 temp_stall_ts = 1;
1039                 if (output_format & FIO_OUTPUT_NORMAL) {
1040                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1041                                  o->name,
1042                                  need_extend > 1 ? "s" : "",
1043                                  need_extend,
1044                                  need_extend > 1 ? "s" : "",
1045                                  need_extend > 1 ? "total " : "",
1046                                  extend_size >> 20);
1047                 }
1048
1049                 for_each_file(td, f, i) {
1050                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1051
1052                         if (!fio_file_extend(f))
1053                                 continue;
1054
1055                         assert(f->filetype == FIO_TYPE_FILE);
1056                         fio_file_clear_extend(f);
1057                         if (!o->fill_device) {
1058                                 old_len = f->real_file_size;
1059                                 extend_len = f->io_size + f->file_offset -
1060                                                 old_len;
1061                         }
1062                         f->real_file_size = (f->io_size + f->file_offset);
1063                         err = extend_file(td, f);
1064                         if (err)
1065                                 break;
1066
1067                         err = __file_invalidate_cache(td, f, old_len,
1068                                                                 extend_len);
1069
1070                         /*
1071                          * Shut up static checker
1072                          */
1073                         if (f->fd != -1)
1074                                 close(f->fd);
1075
1076                         f->fd = -1;
1077                         if (err)
1078                                 break;
1079                 }
1080                 temp_stall_ts = 0;
1081         }
1082
1083         if (err)
1084                 goto err_out;
1085
1086         if (!o->zone_size)
1087                 o->zone_size = o->size;
1088
1089         /*
1090          * iolog already set the total io size, if we read back
1091          * stored entries.
1092          */
1093         if (!o->read_iolog_file) {
1094                 if (o->io_size)
1095                         td->total_io_size = o->io_size * o->loops;
1096                 else
1097                         td->total_io_size = o->size * o->loops;
1098         }
1099
1100 done:
1101         if (o->create_only)
1102                 td->done = 1;
1103
1104         td_restore_runstate(td, old_state);
1105         return 0;
1106 err_offset:
1107         log_err("%s: you need to specify valid offset=\n", o->name);
1108 err_out:
1109         td_restore_runstate(td, old_state);
1110         return 1;
1111 }
1112
1113 int pre_read_files(struct thread_data *td)
1114 {
1115         struct fio_file *f;
1116         unsigned int i;
1117
1118         dprint(FD_FILE, "pre_read files\n");
1119
1120         for_each_file(td, f, i) {
1121                 pre_read_file(td, f);
1122         }
1123
1124         return 1;
1125 }
1126
1127 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1128 {
1129         unsigned int range_size, seed;
1130         unsigned long nranges;
1131         uint64_t fsize;
1132
1133         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1134         fsize = min(f->real_file_size, f->io_size);
1135
1136         nranges = (fsize + range_size - 1) / range_size;
1137
1138         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1139         if (!td->o.rand_repeatable)
1140                 seed = td->rand_seeds[4];
1141
1142         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1143                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1144         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1145                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1146         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1147                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1148
1149         return 1;
1150 }
1151
1152 static int init_rand_distribution(struct thread_data *td)
1153 {
1154         struct fio_file *f;
1155         unsigned int i;
1156         int state;
1157
1158         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1159                 return 0;
1160
1161         state = td_bump_runstate(td, TD_SETTING_UP);
1162
1163         for_each_file(td, f, i)
1164                 __init_rand_distribution(td, f);
1165
1166         td_restore_runstate(td, state);
1167
1168         return 1;
1169 }
1170
1171 /*
1172  * Check if the number of blocks exceeds the randomness capability of
1173  * the selected generator. Tausworthe is 32-bit, the others are fullly
1174  * 64-bit capable.
1175  */
1176 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1177                                  uint64_t blocks)
1178 {
1179         if (blocks <= FRAND32_MAX)
1180                 return 0;
1181         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1182                 return 0;
1183
1184         /*
1185          * If the user hasn't specified a random generator, switch
1186          * to tausworthe64 with informational warning. If the user did
1187          * specify one, just warn.
1188          */
1189         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1190                         f->file_name);
1191
1192         if (!fio_option_is_set(&td->o, random_generator)) {
1193                 log_info("fio: Switching to tausworthe64. Use the "
1194                          "random_generator= option to get rid of this "
1195                          "warning.\n");
1196                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1197                 return 0;
1198         }
1199
1200         /*
1201          * Just make this information to avoid breaking scripts.
1202          */
1203         log_info("fio: Use the random_generator= option to switch to lfsr or "
1204                          "tausworthe64.\n");
1205         return 0;
1206 }
1207
1208 int init_random_map(struct thread_data *td)
1209 {
1210         unsigned long long blocks;
1211         struct fio_file *f;
1212         unsigned int i;
1213
1214         if (init_rand_distribution(td))
1215                 return 0;
1216         if (!td_random(td))
1217                 return 0;
1218
1219         for_each_file(td, f, i) {
1220                 uint64_t fsize = min(f->real_file_size, f->io_size);
1221
1222                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1223
1224                 if (check_rand_gen_limits(td, f, blocks))
1225                         return 1;
1226
1227                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1228                         unsigned long seed;
1229
1230                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1231
1232                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1233                                 fio_file_set_lfsr(f);
1234                                 continue;
1235                         }
1236                 } else if (!td->o.norandommap) {
1237                         f->io_axmap = axmap_new(blocks);
1238                         if (f->io_axmap) {
1239                                 fio_file_set_axmap(f);
1240                                 continue;
1241                         }
1242                 } else if (td->o.norandommap)
1243                         continue;
1244
1245                 if (!td->o.softrandommap) {
1246                         log_err("fio: failed allocating random map. If running"
1247                                 " a large number of jobs, try the 'norandommap'"
1248                                 " option or set 'softrandommap'. Or give"
1249                                 " a larger --alloc-size to fio.\n");
1250                         return 1;
1251                 }
1252
1253                 log_info("fio: file %s failed allocating random map. Running "
1254                          "job without.\n", f->file_name);
1255         }
1256
1257         return 0;
1258 }
1259
1260 void close_files(struct thread_data *td)
1261 {
1262         struct fio_file *f;
1263         unsigned int i;
1264
1265         for_each_file(td, f, i) {
1266                 if (fio_file_open(f))
1267                         td_io_close_file(td, f);
1268         }
1269 }
1270
1271 void close_and_free_files(struct thread_data *td)
1272 {
1273         struct fio_file *f;
1274         unsigned int i;
1275
1276         dprint(FD_FILE, "close files\n");
1277
1278         for_each_file(td, f, i) {
1279                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1280                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1281                         td_io_unlink_file(td, f);
1282                 }
1283
1284                 if (fio_file_open(f))
1285                         td_io_close_file(td, f);
1286
1287                 remove_file_hash(f);
1288
1289                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1290                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1291                         td_io_unlink_file(td, f);
1292                 }
1293
1294                 sfree(f->file_name);
1295                 f->file_name = NULL;
1296                 if (fio_file_axmap(f)) {
1297                         axmap_free(f->io_axmap);
1298                         f->io_axmap = NULL;
1299                 }
1300                 sfree(f);
1301         }
1302
1303         td->o.filename = NULL;
1304         free(td->files);
1305         free(td->file_locks);
1306         td->files_index = 0;
1307         td->files = NULL;
1308         td->file_locks = NULL;
1309         td->o.file_lock_mode = FILE_LOCK_NONE;
1310         td->o.nr_files = 0;
1311 }
1312
1313 static void get_file_type(struct fio_file *f)
1314 {
1315         struct stat sb;
1316
1317         if (!strcmp(f->file_name, "-"))
1318                 f->filetype = FIO_TYPE_PIPE;
1319         else
1320                 f->filetype = FIO_TYPE_FILE;
1321
1322 #ifdef WIN32
1323         /* \\.\ is the device namespace in Windows, where every file is
1324          * a block device */
1325         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1326                 f->filetype = FIO_TYPE_BLOCK;
1327 #endif
1328
1329         if (!stat(f->file_name, &sb)) {
1330                 if (S_ISBLK(sb.st_mode))
1331                         f->filetype = FIO_TYPE_BLOCK;
1332                 else if (S_ISCHR(sb.st_mode))
1333                         f->filetype = FIO_TYPE_CHAR;
1334                 else if (S_ISFIFO(sb.st_mode))
1335                         f->filetype = FIO_TYPE_PIPE;
1336         }
1337 }
1338
1339 static bool __is_already_allocated(const char *fname, bool set)
1340 {
1341         struct flist_head *entry;
1342         bool ret;
1343
1344         ret = file_bloom_exists(fname, set);
1345         if (!ret)
1346                 return ret;
1347
1348         flist_for_each(entry, &filename_list) {
1349                 struct file_name *fn;
1350
1351                 fn = flist_entry(entry, struct file_name, list);
1352
1353                 if (!strcmp(fn->filename, fname))
1354                         return true;
1355         }
1356
1357         return false;
1358 }
1359
1360 static bool is_already_allocated(const char *fname)
1361 {
1362         bool ret;
1363
1364         fio_file_hash_lock();
1365         ret = __is_already_allocated(fname, false);
1366         fio_file_hash_unlock();
1367
1368         return ret;
1369 }
1370
1371 static void set_already_allocated(const char *fname)
1372 {
1373         struct file_name *fn;
1374
1375         fn = malloc(sizeof(struct file_name));
1376         fn->filename = strdup(fname);
1377
1378         fio_file_hash_lock();
1379         if (!__is_already_allocated(fname, true)) {
1380                 flist_add_tail(&fn->list, &filename_list);
1381                 fn = NULL;
1382         }
1383         fio_file_hash_unlock();
1384
1385         if (fn) {
1386                 free(fn->filename);
1387                 free(fn);
1388         }
1389 }
1390
1391 static void free_already_allocated(void)
1392 {
1393         struct flist_head *entry, *tmp;
1394         struct file_name *fn;
1395
1396         if (flist_empty(&filename_list))
1397                 return;
1398
1399         fio_file_hash_lock();
1400         flist_for_each_safe(entry, tmp, &filename_list) {
1401                 fn = flist_entry(entry, struct file_name, list);
1402                 free(fn->filename);
1403                 flist_del(&fn->list);
1404                 free(fn);
1405         }
1406
1407         fio_file_hash_unlock();
1408 }
1409
1410 static struct fio_file *alloc_new_file(struct thread_data *td)
1411 {
1412         struct fio_file *f;
1413
1414         f = smalloc(sizeof(*f));
1415         if (!f) {
1416                 assert(0);
1417                 return NULL;
1418         }
1419
1420         f->fd = -1;
1421         f->shadow_fd = -1;
1422         fio_file_reset(td, f);
1423         return f;
1424 }
1425
1426 bool exists_and_not_regfile(const char *filename)
1427 {
1428         struct stat sb;
1429
1430         if (lstat(filename, &sb) == -1)
1431                 return false;
1432
1433 #ifndef WIN32 /* NOT Windows */
1434         if (S_ISREG(sb.st_mode))
1435                 return false;
1436 #else
1437         /* \\.\ is the device namespace in Windows, where every file
1438          * is a device node */
1439         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1440                 return false;
1441 #endif
1442
1443         return true;
1444 }
1445
1446 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1447 {
1448         int cur_files = td->files_index;
1449         char file_name[PATH_MAX];
1450         struct fio_file *f;
1451         int len = 0;
1452
1453         dprint(FD_FILE, "add file %s\n", fname);
1454
1455         if (td->o.directory)
1456                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1457                                         td->o.unique_filename);
1458
1459         sprintf(file_name + len, "%s", fname);
1460
1461         /* clean cloned siblings using existing files */
1462         if (numjob && is_already_allocated(file_name) &&
1463             !exists_and_not_regfile(fname))
1464                 return 0;
1465
1466         f = alloc_new_file(td);
1467
1468         if (td->files_size <= td->files_index) {
1469                 unsigned int new_size = td->o.nr_files + 1;
1470
1471                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1472
1473                 td->files = realloc(td->files, new_size * sizeof(f));
1474                 if (td->files == NULL) {
1475                         log_err("fio: realloc OOM\n");
1476                         assert(0);
1477                 }
1478                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1479                         td->file_locks = realloc(td->file_locks, new_size);
1480                         if (!td->file_locks) {
1481                                 log_err("fio: realloc OOM\n");
1482                                 assert(0);
1483                         }
1484                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1485                 }
1486                 td->files_size = new_size;
1487         }
1488         td->files[cur_files] = f;
1489         f->fileno = cur_files;
1490
1491         /*
1492          * init function, io engine may not be loaded yet
1493          */
1494         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1495                 f->real_file_size = -1ULL;
1496
1497         f->file_name = smalloc_strdup(file_name);
1498         if (!f->file_name)
1499                 assert(0);
1500
1501         get_file_type(f);
1502
1503         switch (td->o.file_lock_mode) {
1504         case FILE_LOCK_NONE:
1505                 break;
1506         case FILE_LOCK_READWRITE:
1507                 f->rwlock = fio_rwlock_init();
1508                 break;
1509         case FILE_LOCK_EXCLUSIVE:
1510                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1511                 break;
1512         default:
1513                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1514                 assert(0);
1515         }
1516
1517         td->files_index++;
1518         if (f->filetype == FIO_TYPE_FILE)
1519                 td->nr_normal_files++;
1520
1521         set_already_allocated(file_name);
1522
1523         if (inc)
1524                 td->o.nr_files++;
1525
1526         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1527                                                         cur_files);
1528
1529         return cur_files;
1530 }
1531
1532 int add_file_exclusive(struct thread_data *td, const char *fname)
1533 {
1534         struct fio_file *f;
1535         unsigned int i;
1536
1537         for_each_file(td, f, i) {
1538                 if (!strcmp(f->file_name, fname))
1539                         return i;
1540         }
1541
1542         return add_file(td, fname, 0, 1);
1543 }
1544
1545 void get_file(struct fio_file *f)
1546 {
1547         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1548         assert(fio_file_open(f));
1549         f->references++;
1550 }
1551
1552 int put_file(struct thread_data *td, struct fio_file *f)
1553 {
1554         int f_ret = 0, ret = 0;
1555
1556         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1557
1558         if (!fio_file_open(f)) {
1559                 assert(f->fd == -1);
1560                 return 0;
1561         }
1562
1563         assert(f->references);
1564         if (--f->references)
1565                 return 0;
1566
1567         if (should_fsync(td) && td->o.fsync_on_close) {
1568                 f_ret = fsync(f->fd);
1569                 if (f_ret < 0)
1570                         f_ret = errno;
1571         }
1572
1573         if (td->io_ops->close_file)
1574                 ret = td->io_ops->close_file(td, f);
1575
1576         if (!ret)
1577                 ret = f_ret;
1578
1579         td->nr_open_files--;
1580         fio_file_clear_open(f);
1581         assert(f->fd == -1);
1582         return ret;
1583 }
1584
1585 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1586 {
1587         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1588                 return;
1589
1590         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1591                 if (ddir == DDIR_READ)
1592                         fio_rwlock_read(f->rwlock);
1593                 else
1594                         fio_rwlock_write(f->rwlock);
1595         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1596                 fio_mutex_down(f->lock);
1597
1598         td->file_locks[f->fileno] = td->o.file_lock_mode;
1599 }
1600
1601 void unlock_file(struct thread_data *td, struct fio_file *f)
1602 {
1603         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1604                 return;
1605
1606         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1607                 fio_rwlock_unlock(f->rwlock);
1608         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1609                 fio_mutex_up(f->lock);
1610
1611         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1612 }
1613
1614 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1615 {
1616         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1617                 return;
1618         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1619                 unlock_file(td, f);
1620 }
1621
1622 static int recurse_dir(struct thread_data *td, const char *dirname)
1623 {
1624         struct dirent *dir;
1625         int ret = 0;
1626         DIR *D;
1627
1628         D = opendir(dirname);
1629         if (!D) {
1630                 char buf[FIO_VERROR_SIZE];
1631
1632                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1633                 td_verror(td, errno, buf);
1634                 return 1;
1635         }
1636
1637         while ((dir = readdir(D)) != NULL) {
1638                 char full_path[PATH_MAX];
1639                 struct stat sb;
1640
1641                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1642                         continue;
1643
1644                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1645
1646                 if (lstat(full_path, &sb) == -1) {
1647                         if (errno != ENOENT) {
1648                                 td_verror(td, errno, "stat");
1649                                 ret = 1;
1650                                 break;
1651                         }
1652                 }
1653
1654                 if (S_ISREG(sb.st_mode)) {
1655                         add_file(td, full_path, 0, 1);
1656                         continue;
1657                 }
1658                 if (!S_ISDIR(sb.st_mode))
1659                         continue;
1660
1661                 ret = recurse_dir(td, full_path);
1662                 if (ret)
1663                         break;
1664         }
1665
1666         closedir(D);
1667         return ret;
1668 }
1669
1670 int add_dir_files(struct thread_data *td, const char *path)
1671 {
1672         int ret = recurse_dir(td, path);
1673
1674         if (!ret)
1675                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1676
1677         return ret;
1678 }
1679
1680 void dup_files(struct thread_data *td, struct thread_data *org)
1681 {
1682         struct fio_file *f;
1683         unsigned int i;
1684
1685         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1686
1687         if (!org->files)
1688                 return;
1689
1690         td->files = malloc(org->files_index * sizeof(f));
1691
1692         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1693                 td->file_locks = malloc(org->files_index);
1694
1695         for_each_file(org, f, i) {
1696                 struct fio_file *__f;
1697
1698                 __f = alloc_new_file(td);
1699
1700                 if (f->file_name) {
1701                         __f->file_name = smalloc_strdup(f->file_name);
1702                         if (!__f->file_name)
1703                                 assert(0);
1704
1705                         __f->filetype = f->filetype;
1706                 }
1707
1708                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1709                         __f->lock = f->lock;
1710                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1711                         __f->rwlock = f->rwlock;
1712
1713                 td->files[i] = __f;
1714         }
1715 }
1716
1717 /*
1718  * Returns the index that matches the filename, or -1 if not there
1719  */
1720 int get_fileno(struct thread_data *td, const char *fname)
1721 {
1722         struct fio_file *f;
1723         unsigned int i;
1724
1725         for_each_file(td, f, i)
1726                 if (!strcmp(f->file_name, fname))
1727                         return i;
1728
1729         return -1;
1730 }
1731
1732 /*
1733  * For log usage, where we add/open/close files automatically
1734  */
1735 void free_release_files(struct thread_data *td)
1736 {
1737         close_files(td);
1738         td->o.nr_files = 0;
1739         td->o.open_files = 0;
1740         td->files_index = 0;
1741         td->nr_normal_files = 0;
1742 }
1743
1744 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1745 {
1746         int i;
1747
1748         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1749                 f->last_pos[i] = f->file_offset;
1750                 f->last_start[i] = -1ULL;
1751         }
1752
1753         if (fio_file_axmap(f))
1754                 axmap_reset(f->io_axmap);
1755         else if (fio_file_lfsr(f))
1756                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1757 }
1758
1759 bool fio_files_done(struct thread_data *td)
1760 {
1761         struct fio_file *f;
1762         unsigned int i;
1763
1764         for_each_file(td, f, i)
1765                 if (!fio_file_done(f))
1766                         return false;
1767
1768         return true;
1769 }
1770
1771 /* free memory used in initialization phase only */
1772 void filesetup_mem_free(void)
1773 {
1774         free_already_allocated();
1775 }