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             td_ioengine_flagged(td, FIO_NOIO))
 240                 return 0;
 241
 242         if (!fio_file_open(f)) {
 243                 if (td->io_ops->open_file(td, f)) {
 244                         log_err("fio: cannot pre-read, failed to open file\n");
 245                         return 1;
 246                 }
 247                 did_open = 1;
 248         }
 249
 250         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 251
 252         left = f->io_size;
 253         bs = td->o.max_bs[DDIR_READ];
 254         if (bs > left)
 255                 bs = left;
 256
 257         b = malloc(bs);
 258         if (!b) {
 259                 td_verror(td, errno, "malloc");
 260                 ret = 1;
 261                 goto error;
 262         }
 263         memset(b, 0, bs);
 264
 265         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 266                 td_verror(td, errno, "lseek");
 267                 log_err("fio: failed to lseek pre-read file\n");
 268                 ret = 1;
 269                 goto error;
 270         }
 271
 272         while (left && !td->terminate) {
 273                 if (bs > left)
 274                         bs = left;
 275
 276                 r = read(f->fd, b, bs);
 277
 278                 if (r == (int) bs) {
 279                         left -= bs;
 280                         continue;
 281                 } else {
 282                         td_verror(td, EIO, "pre_read");
 283                         break;
 284                 }
 285         }
 286
 287 error:
 288         td_restore_runstate(td, old_runstate);
 289
 290         if (did_open)
 291                 td->io_ops->close_file(td, f);
 292
 293         free(b);
 294         return ret;
 295 }
 296
 297 unsigned long long get_rand_file_size(struct thread_data *td)
 298 {
 299         unsigned long long ret, sized;
 300         uint64_t frand_max;
 301         unsigned long r;
 302
 303         frand_max = rand_max(&td->file_size_state);
 304         r = __rand(&td->file_size_state);
 305         sized = td->o.file_size_high - td->o.file_size_low;
 306         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 307         ret += td->o.file_size_low;
 308         ret -= (ret % td->o.rw_min_bs);
 309         return ret;
 310 }
 311
 312 static int file_size(struct thread_data *td, struct fio_file *f)
 313 {
 314         struct stat st;
 315
 316         if (stat(f->file_name, &st) == -1) {
 317                 td_verror(td, errno, "fstat");
 318                 return 1;
 319         }
 320
 321         f->real_file_size = st.st_size;
 322         return 0;
 323 }
 324
 325 static int bdev_size(struct thread_data *td, struct fio_file *f)
 326 {
 327         unsigned long long bytes = 0;
 328         int r;
 329
 330         if (td->io_ops->open_file(td, f)) {
 331                 log_err("fio: failed opening blockdev %s for size check\n",
 332                         f->file_name);
 333                 return 1;
 334         }
 335
 336         r = blockdev_size(f, &bytes);
 337         if (r) {
 338                 td_verror(td, r, "blockdev_size");
 339                 goto err;
 340         }
 341
 342         if (!bytes) {
 343                 log_err("%s: zero sized block device?\n", f->file_name);
 344                 goto err;
 345         }
 346
 347         f->real_file_size = bytes;
 348         td->io_ops->close_file(td, f);
 349         return 0;
 350 err:
 351         td->io_ops->close_file(td, f);
 352         return 1;
 353 }
 354
 355 static int char_size(struct thread_data *td, struct fio_file *f)
 356 {
 357 #ifdef FIO_HAVE_CHARDEV_SIZE
 358         unsigned long long bytes = 0;
 359         int r;
 360
 361         if (td->io_ops->open_file(td, f)) {
 362                 log_err("fio: failed opening chardev %s for size check\n",
 363                         f->file_name);
 364                 return 1;
 365         }
 366
 367         r = chardev_size(f, &bytes);
 368         if (r) {
 369                 td_verror(td, r, "chardev_size");
 370                 goto err;
 371         }
 372
 373         if (!bytes) {
 374                 log_err("%s: zero sized char device?\n", f->file_name);
 375                 goto err;
 376         }
 377
 378         f->real_file_size = bytes;
 379         td->io_ops->close_file(td, f);
 380         return 0;
 381 err:
 382         td->io_ops->close_file(td, f);
 383         return 1;
 384 #else
 385         f->real_file_size = -1ULL;
 386         return 0;
 387 #endif
 388 }
 389
 390 static int get_file_size(struct thread_data *td, struct fio_file *f)
 391 {
 392         int ret = 0;
 393
 394         if (fio_file_size_known(f))
 395                 return 0;
 396
 397         if (f->filetype == FIO_TYPE_FILE)
 398                 ret = file_size(td, f);
 399         else if (f->filetype == FIO_TYPE_BLOCK)
 400                 ret = bdev_size(td, f);
 401         else if (f->filetype == FIO_TYPE_CHAR)
 402                 ret = char_size(td, f);
 403         else
 404                 f->real_file_size = -1ULL;
 405
 406         /*
 407          * Leave ->real_file_size with 0 since it could be expectation
 408          * of initial setup for regular files.
 409          */
 410         if (ret)
 411                 return ret;
 412
 413         /*
 414          * If ->real_file_size is -1, a conditional for the message
 415          * "offset extends end" is always true, but it makes no sense,
 416          * so just return the same value here.
 417          */
 418         if (f->real_file_size == -1ULL) {
 419                 log_info("%s: failed to get file size of %s\n", td->o.name,
 420                                         f->file_name);
 421                 return 1;
 422         }
 423
 424         if (td->o.start_offset && f->file_offset == 0)
 425                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 426                                         f->file_name);
 427         /*
 428          * ->file_offset normally hasn't been initialized yet, so this
 429          * is basically always false.
 430          */
 431         if (f->file_offset > f->real_file_size) {
 432                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 433                                         (unsigned long long) f->file_offset,
 434                                         (unsigned long long) f->real_file_size);
 435                 return 1;
 436         }
 437
 438         fio_file_set_size_known(f);
 439         return 0;
 440 }
 441
 442 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 443                                    unsigned long long off,
 444                                    unsigned long long len)
 445 {
 446         int errval = 0, ret = 0;
 447
 448 #ifdef CONFIG_ESX
 449         return 0;
 450 #endif
 451
 452         if (len == -1ULL)
 453                 len = f->io_size;
 454         if (off == -1ULL)
 455                 off = f->file_offset;
 456
 457         if (len == -1ULL || off == -1ULL)
 458                 return 0;
 459
 460         if (td->io_ops->invalidate) {
 461                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 462                         f->file_name);
 463                 ret = td->io_ops->invalidate(td, f);
 464                 if (ret < 0)
 465                         errval = -ret;
 466         } else if (f->filetype == FIO_TYPE_FILE) {
 467                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 468                         f->file_name, off, len);
 469                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 470                 if (ret)
 471                         errval = ret;
 472         } else if (f->filetype == FIO_TYPE_BLOCK) {
 473                 int retry_count = 0;
 474
 475                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 476                 ret = blockdev_invalidate_cache(f);
 477                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 478                         /*
 479                          * Linux multipath devices reject ioctl while
 480                          * the maps are being updated. That window can
 481                          * last tens of milliseconds; we'll try up to
 482                          * a quarter of a second.
 483                          */
 484                         usleep(10000);
 485                         ret = blockdev_invalidate_cache(f);
 486                 }
 487                 if (ret < 0 && errno == EACCES && geteuid()) {
 488                         if (!root_warn) {
 489                                 log_err("fio: only root may flush block "
 490                                         "devices. Cache flush bypassed!\n");
 491                                 root_warn = 1;
 492                         }
 493                         ret = 0;
 494                 }
 495                 if (ret < 0)
 496                         errval = errno;
 497                 else if (ret) /* probably not supported */
 498                         errval = ret;
 499         } else if (f->filetype == FIO_TYPE_CHAR ||
 500                    f->filetype == FIO_TYPE_PIPE) {
 501                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 502                 ret = 0;
 503         }
 504
 505         /*
 506          * Cache flushing isn't a fatal condition, and we know it will
 507          * happen on some platforms where we don't have the proper
 508          * function to flush eg block device caches. So just warn and
 509          * continue on our way.
 510          */
 511         if (errval)
 512                 log_info("fio: cache invalidation of %s failed: %s\n",
 513                          f->file_name, strerror(errval));
 514
 515         return 0;
 516
 517 }
 518
 519 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 520 {
 521         if (!fio_file_open(f))
 522                 return 0;
 523
 524         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 525 }
 526
 527 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 528 {
 529         int ret = 0;
 530
 531         dprint(FD_FILE, "fd close %s\n", f->file_name);
 532
 533         remove_file_hash(f);
 534
 535         if (close(f->fd) < 0)
 536                 ret = errno;
 537
 538         f->fd = -1;
 539
 540         if (f->shadow_fd != -1) {
 541                 close(f->shadow_fd);
 542                 f->shadow_fd = -1;
 543         }
 544
 545         f->engine_pos = 0;
 546         return ret;
 547 }
 548
 549 int file_lookup_open(struct fio_file *f, int flags)
 550 {
 551         struct fio_file *__f;
 552         int from_hash;
 553
 554         __f = lookup_file_hash(f->file_name);
 555         if (__f) {
 556                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 557                 f->lock = __f->lock;
 558                 from_hash = 1;
 559         } else {
 560                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 561                 from_hash = 0;
 562         }
 563
 564 #ifdef WIN32
 565         flags |= _O_BINARY;
 566 #endif
 567
 568         f->fd = open(f->file_name, flags, 0600);
 569         return from_hash;
 570 }
 571
 572 static int file_close_shadow_fds(struct thread_data *td)
 573 {
 574         struct fio_file *f;
 575         int num_closed = 0;
 576         unsigned int i;
 577
 578         for_each_file(td, f, i) {
 579                 if (f->shadow_fd == -1)
 580                         continue;
 581
 582                 close(f->shadow_fd);
 583                 f->shadow_fd = -1;
 584                 num_closed++;
 585         }
 586
 587         return num_closed;
 588 }
 589
 590 int generic_open_file(struct thread_data *td, struct fio_file *f)
 591 {
 592         int is_std = 0;
 593         int flags = 0;
 594         int from_hash = 0;
 595
 596         dprint(FD_FILE, "fd open %s\n", f->file_name);
 597
 598         if (!strcmp(f->file_name, "-")) {
 599                 if (td_rw(td)) {
 600                         log_err("fio: can't read/write to stdin/out\n");
 601                         return 1;
 602                 }
 603                 is_std = 1;
 604
 605                 /*
 606                  * move output logging to stderr, if we are writing to stdout
 607                  */
 608                 if (td_write(td))
 609                         f_out = stderr;
 610         }
 611
 612         if (td_trim(td))
 613                 goto skip_flags;
 614         if (td->o.odirect)
 615                 flags |= OS_O_DIRECT;
 616         if (td->o.oatomic) {
 617                 if (!FIO_O_ATOMIC) {
 618                         td_verror(td, EINVAL, "OS does not support atomic IO");
 619                         return 1;
 620                 }
 621                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 622         }
 623         if (td->o.sync_io)
 624                 flags |= O_SYNC;
 625         if (td->o.create_on_open && td->o.allow_create)
 626                 flags |= O_CREAT;
 627 skip_flags:
 628         if (f->filetype != FIO_TYPE_FILE)
 629                 flags |= FIO_O_NOATIME;
 630
 631 open_again:
 632         if (td_write(td)) {
 633                 if (!read_only)
 634                         flags |= O_RDWR;
 635
 636                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 637                         flags |= O_CREAT;
 638
 639                 if (is_std)
 640                         f->fd = dup(STDOUT_FILENO);
 641                 else
 642                         from_hash = file_lookup_open(f, flags);
 643         } else if (td_read(td)) {
 644                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 645                         flags |= O_RDWR;
 646                 else
 647                         flags |= O_RDONLY;
 648
 649                 if (is_std)
 650                         f->fd = dup(STDIN_FILENO);
 651                 else
 652                         from_hash = file_lookup_open(f, flags);
 653         } else if (td_trim(td)) {
 654                 assert(!td_rw(td)); /* should have matched above */
 655                 flags |= O_RDWR;
 656                 from_hash = file_lookup_open(f, flags);
 657         }
 658
 659         if (f->fd == -1) {
 660                 char buf[FIO_VERROR_SIZE];
 661                 int __e = errno;
 662
 663                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 664                         flags &= ~FIO_O_NOATIME;
 665                         goto open_again;
 666                 }
 667                 if (__e == EMFILE && file_close_shadow_fds(td))
 668                         goto open_again;
 669
 670                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 671
 672                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 673                         log_err("fio: looks like your file system does not " \
 674                                 "support direct=1/buffered=0\n");
 675                 }
 676
 677                 td_verror(td, __e, buf);
 678                 return 1;
 679         }
 680
 681         if (!from_hash && f->fd != -1) {
 682                 if (add_file_hash(f)) {
 683                         int fio_unused ret;
 684
 685                         /*
 686                          * Stash away descriptor for later close. This is to
 687                          * work-around a "feature" on Linux, where a close of
 688                          * an fd that has been opened for write will trigger
 689                          * udev to call blkid to check partitions, fs id, etc.
 690                          * That pollutes the device cache, which can slow down
 691                          * unbuffered accesses.
 692                          */
 693                         if (f->shadow_fd == -1)
 694                                 f->shadow_fd = f->fd;
 695                         else {
 696                                 /*
 697                                  * OK to ignore, we haven't done anything
 698                                  * with it
 699                                  */
 700                                 ret = generic_close_file(td, f);
 701                         }
 702                         goto open_again;
 703                 }
 704         }
 705
 706         return 0;
 707 }
 708
 709 /*
 710  * This function i.e. get_file_size() is the default .get_file_size
 711  * implementation of majority of I/O engines.
 712  */
 713 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 714 {
 715         return get_file_size(td, f);
 716 }
 717
 718 /*
 719  * open/close all files, so that ->real_file_size gets set
 720  */
 721 static int get_file_sizes(struct thread_data *td)
 722 {
 723         struct fio_file *f;
 724         unsigned int i;
 725         int err = 0;
 726
 727         for_each_file(td, f, i) {
 728                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 729                                                                 f->file_name);
 730
 731                 if (td_io_get_file_size(td, f)) {
 732                         if (td->error != ENOENT) {
 733                                 log_err("%s\n", td->verror);
 734                                 err = 1;
 735                                 break;
 736                         }
 737                         clear_error(td);
 738                 }
 739
 740                 /*
 741                  * There are corner cases where we end up with -1 for
 742                  * ->real_file_size due to unsupported file type, etc.
 743                  * We then just set to size option value divided by number
 744                  * of files, similar to the way file ->io_size is set.
 745                  * stat(2) failure doesn't set ->real_file_size to -1.
 746                  */
 747                 if (f->real_file_size == -1ULL && td->o.size)
 748                         f->real_file_size = td->o.size / td->o.nr_files;
 749         }
 750
 751         return err;
 752 }
 753
 754 struct fio_mount {
 755         struct flist_head list;
 756         const char *base;
 757         char __base[256];
 758         unsigned int key;
 759 };
 760
 761 /*
 762  * Get free number of bytes for each file on each unique mount.
 763  */
 764 static unsigned long long get_fs_free_counts(struct thread_data *td)
 765 {
 766         struct flist_head *n, *tmp;
 767         unsigned long long ret = 0;
 768         struct fio_mount *fm;
 769         FLIST_HEAD(list);
 770         struct fio_file *f;
 771         unsigned int i;
 772
 773         for_each_file(td, f, i) {
 774                 struct stat sb;
 775                 char buf[256];
 776
 777                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 778                         if (f->real_file_size != -1ULL)
 779                                 ret += f->real_file_size;
 780                         continue;
 781                 } else if (f->filetype != FIO_TYPE_FILE)
 782                         continue;
 783
 784                 buf[255] = '\0';
 785                 strncpy(buf, f->file_name, 255);
 786
 787                 if (stat(buf, &sb) < 0) {
 788                         if (errno != ENOENT)
 789                                 break;
 790                         strcpy(buf, ".");
 791                         if (stat(buf, &sb) < 0)
 792                                 break;
 793                 }
 794
 795                 fm = NULL;
 796                 flist_for_each(n, &list) {
 797                         fm = flist_entry(n, struct fio_mount, list);
 798                         if (fm->key == sb.st_dev)
 799                                 break;
 800
 801                         fm = NULL;
 802                 }
 803
 804                 if (fm)
 805                         continue;
 806
 807                 fm = calloc(1, sizeof(*fm));
 808                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 809                 fm->base = basename(fm->__base);
 810                 fm->key = sb.st_dev;
 811                 flist_add(&fm->list, &list);
 812         }
 813
 814         flist_for_each_safe(n, tmp, &list) {
 815                 unsigned long long sz;
 816
 817                 fm = flist_entry(n, struct fio_mount, list);
 818                 flist_del(&fm->list);
 819
 820                 sz = get_fs_free_size(fm->base);
 821                 if (sz && sz != -1ULL)
 822                         ret += sz;
 823
 824                 free(fm);
 825         }
 826
 827         return ret;
 828 }
 829
 830 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 831 {
 832         struct thread_options *o = &td->o;
 833
 834         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 835                 return f->real_file_size;
 836
 837         return td->o.start_offset +
 838                 td->subjob_number * td->o.offset_increment;
 839 }
 840
 841 /*
 842  * Open the files and setup files sizes, creating files if necessary.
 843  */
 844 int setup_files(struct thread_data *td)
 845 {
 846         unsigned long long total_size, extend_size;
 847         struct thread_options *o = &td->o;
 848         struct fio_file *f;
 849         unsigned int i, nr_fs_extra = 0;
 850         int err = 0, need_extend;
 851         int old_state;
 852         const unsigned int bs = td_min_bs(td);
 853         uint64_t fs = 0;
 854
 855         dprint(FD_FILE, "setup files\n");
 856
 857         old_state = td_bump_runstate(td, TD_SETTING_UP);
 858
 859         if (o->read_iolog_file)
 860                 goto done;
 861
 862         /*
 863          * Find out physical size of files or devices for this thread,
 864          * before we determine I/O size and range of our targets.
 865          * If ioengine defines a setup() method, it's responsible for
 866          * opening the files and setting f->real_file_size to indicate
 867          * the valid range for that file.
 868          */
 869         if (td->io_ops->setup)
 870                 err = td->io_ops->setup(td);
 871         else
 872                 err = get_file_sizes(td);
 873
 874         if (err)
 875                 goto err_out;
 876
 877         /*
 878          * check sizes. if the files/devices do not exist and the size
 879          * isn't passed to fio, abort.
 880          */
 881         total_size = 0;
 882         for_each_file(td, f, i) {
 883                 f->fileno = i;
 884                 if (f->real_file_size == -1ULL)
 885                         total_size = -1ULL;
 886                 else
 887                         total_size += f->real_file_size;
 888         }
 889
 890         if (o->fill_device)
 891                 td->fill_device_size = get_fs_free_counts(td);
 892
 893         /*
 894          * device/file sizes are zero and no size given, punt
 895          */
 896         if ((!total_size || total_size == -1ULL) && !o->size &&
 897             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 898             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 899                 log_err("%s: you need to specify size=\n", o->name);
 900                 td_verror(td, EINVAL, "total_file_size");
 901                 goto err_out;
 902         }
 903
 904         /*
 905          * Calculate per-file size and potential extra size for the
 906          * first files, if needed (i.e. if we don't have a fixed size).
 907          */
 908         if (!o->file_size_low && o->nr_files) {
 909                 uint64_t all_fs;
 910
 911                 fs = o->size / o->nr_files;
 912                 all_fs = fs * o->nr_files;
 913
 914                 if (all_fs < o->size)
 915                         nr_fs_extra = (o->size - all_fs) / bs;
 916         }
 917
 918         /*
 919          * now file sizes are known, so we can set ->io_size. if size= is
 920          * not given, ->io_size is just equal to ->real_file_size. if size
 921          * is given, ->io_size is size / nr_files.
 922          */
 923         extend_size = total_size = 0;
 924         need_extend = 0;
 925         for_each_file(td, f, i) {
 926                 f->file_offset = get_start_offset(td, f);
 927
 928                 /*
 929                  * Update ->io_size depending on options specified.
 930                  * ->file_size_low being 0 means filesize option isn't set.
 931                  * Non zero ->file_size_low equals ->file_size_high means
 932                  * filesize option is set in a fixed size format.
 933                  * Non zero ->file_size_low not equals ->file_size_high means
 934                  * filesize option is set in a range format.
 935                  */
 936                 if (!o->file_size_low) {
 937                         /*
 938                          * no file size or range given, file size is equal to
 939                          * total size divided by number of files. If the size
 940                          * doesn't divide nicely with the min blocksize,
 941                          * make the first files bigger.
 942                          */
 943                         f->io_size = fs;
 944                         if (nr_fs_extra) {
 945                                 nr_fs_extra--;
 946                                 f->io_size += bs;
 947                         }
 948
 949                         /*
 950                          * We normally don't come here for regular files, but
 951                          * if the result is 0 for a regular file, set it to the
 952                          * real file size. This could be size of the existing
 953                          * one if it already exists, but otherwise will be set
 954                          * to 0. A new file won't be created because
 955                          * ->io_size + ->file_offset equals ->real_file_size.
 956                          */
 957                         if (!f->io_size) {
 958                                 if (f->file_offset > f->real_file_size)
 959                                         goto err_offset;
 960                                 f->io_size = f->real_file_size - f->file_offset;
 961                                 if (!f->io_size)
 962                                         log_info("fio: file %s may be ignored\n",
 963                                                 f->file_name);
 964                         }
 965                 } else if (f->real_file_size < o->file_size_low ||
 966                            f->real_file_size > o->file_size_high) {
 967                         if (f->file_offset > o->file_size_low)
 968                                 goto err_offset;
 969                         /*
 970                          * file size given. if it's fixed, use that. if it's a
 971                          * range, generate a random size in-between.
 972                          */
 973                         if (o->file_size_low == o->file_size_high)
 974                                 f->io_size = o->file_size_low - f->file_offset;
 975                         else {
 976                                 f->io_size = get_rand_file_size(td)
 977                                                 - f->file_offset;
 978                         }
 979                 } else
 980                         f->io_size = f->real_file_size - f->file_offset;
 981
 982                 if (f->io_size == -1ULL)
 983                         total_size = -1ULL;
 984                 else {
 985                         if (o->size_percent) {
 986                                 f->io_size = (f->io_size * o->size_percent) / 100;
 987                                 f->io_size -= (f->io_size % td_min_bs(td));
 988                         }
 989                         total_size += f->io_size;
 990                 }
 991
 992                 if (f->filetype == FIO_TYPE_FILE &&
 993                     (f->io_size + f->file_offset) > f->real_file_size &&
 994                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 995                         if (!o->create_on_open) {
 996                                 need_extend++;
 997                                 extend_size += (f->io_size + f->file_offset);
 998                                 fio_file_set_extend(f);
 999                         } else
1000                                 f->real_file_size = f->io_size + f->file_offset;
1001                 }
1002         }
1003
1004         if (td->o.block_error_hist) {
1005                 int len;
1006
1007                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1008                 f = td->files[0];
1009                 len = f->io_size / td->o.bs[DDIR_TRIM];
1010                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1011                         log_err("fio: cannot calculate block histogram with "
1012                                 "%d trim blocks, maximum %d\n",
1013                                 len, MAX_NR_BLOCK_INFOS);
1014                         td_verror(td, EINVAL, "block_error_hist");
1015                         goto err_out;
1016                 }
1017
1018                 td->ts.nr_block_infos = len;
1019                 for (i = 0; i < len; i++)
1020                         td->ts.block_infos[i] =
1021                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1022         } else
1023                 td->ts.nr_block_infos = 0;
1024
1025         if (!o->size || (total_size && o->size > total_size))
1026                 o->size = total_size;
1027
1028         if (o->size < td_min_bs(td)) {
1029                 log_err("fio: blocksize too large for data set\n");
1030                 goto err_out;
1031         }
1032
1033         /*
1034          * See if we need to extend some files, typically needed when our
1035          * target regular files don't exist yet, but our jobs require them
1036          * initially due to read I/Os.
1037          */
1038         if (need_extend) {
1039                 temp_stall_ts = 1;
1040                 if (output_format & FIO_OUTPUT_NORMAL) {
1041                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1042                                  o->name,
1043                                  need_extend > 1 ? "s" : "",
1044                                  need_extend,
1045                                  need_extend > 1 ? "s" : "",
1046                                  need_extend > 1 ? "total " : "",
1047                                  extend_size >> 20);
1048                 }
1049
1050                 for_each_file(td, f, i) {
1051                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1052
1053                         if (!fio_file_extend(f))
1054                                 continue;
1055
1056                         assert(f->filetype == FIO_TYPE_FILE);
1057                         fio_file_clear_extend(f);
1058                         if (!o->fill_device) {
1059                                 old_len = f->real_file_size;
1060                                 extend_len = f->io_size + f->file_offset -
1061                                                 old_len;
1062                         }
1063                         f->real_file_size = (f->io_size + f->file_offset);
1064                         err = extend_file(td, f);
1065                         if (err)
1066                                 break;
1067
1068                         err = __file_invalidate_cache(td, f, old_len,
1069                                                                 extend_len);
1070
1071                         /*
1072                          * Shut up static checker
1073                          */
1074                         if (f->fd != -1)
1075                                 close(f->fd);
1076
1077                         f->fd = -1;
1078                         if (err)
1079                                 break;
1080                 }
1081                 temp_stall_ts = 0;
1082         }
1083
1084         if (err)
1085                 goto err_out;
1086
1087         if (!o->zone_size)
1088                 o->zone_size = o->size;
1089
1090         /*
1091          * iolog already set the total io size, if we read back
1092          * stored entries.
1093          */
1094         if (!o->read_iolog_file) {
1095                 if (o->io_size)
1096                         td->total_io_size = o->io_size * o->loops;
1097                 else
1098                         td->total_io_size = o->size * o->loops;
1099         }
1100
1101 done:
1102         if (o->create_only)
1103                 td->done = 1;
1104
1105         td_restore_runstate(td, old_state);
1106         return 0;
1107 err_offset:
1108         log_err("%s: you need to specify valid offset=\n", o->name);
1109 err_out:
1110         td_restore_runstate(td, old_state);
1111         return 1;
1112 }
1113
1114 int pre_read_files(struct thread_data *td)
1115 {
1116         struct fio_file *f;
1117         unsigned int i;
1118
1119         dprint(FD_FILE, "pre_read files\n");
1120
1121         for_each_file(td, f, i) {
1122                 if (pre_read_file(td, f))
1123                         return -1;
1124         }
1125
1126         return 0;
1127 }
1128
1129 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1130 {
1131         unsigned int range_size, seed;
1132         unsigned long nranges;
1133         uint64_t fsize;
1134
1135         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1136         fsize = min(f->real_file_size, f->io_size);
1137
1138         nranges = (fsize + range_size - 1) / range_size;
1139
1140         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1141         if (!td->o.rand_repeatable)
1142                 seed = td->rand_seeds[4];
1143
1144         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1145                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1146         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1147                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1148         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1149                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1150
1151         return 1;
1152 }
1153
1154 static int init_rand_distribution(struct thread_data *td)
1155 {
1156         struct fio_file *f;
1157         unsigned int i;
1158         int state;
1159
1160         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1161                 return 0;
1162
1163         state = td_bump_runstate(td, TD_SETTING_UP);
1164
1165         for_each_file(td, f, i)
1166                 __init_rand_distribution(td, f);
1167
1168         td_restore_runstate(td, state);
1169
1170         return 1;
1171 }
1172
1173 /*
1174  * Check if the number of blocks exceeds the randomness capability of
1175  * the selected generator. Tausworthe is 32-bit, the others are fullly
1176  * 64-bit capable.
1177  */
1178 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1179                                  uint64_t blocks)
1180 {
1181         if (blocks <= FRAND32_MAX)
1182                 return 0;
1183         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1184                 return 0;
1185
1186         /*
1187          * If the user hasn't specified a random generator, switch
1188          * to tausworthe64 with informational warning. If the user did
1189          * specify one, just warn.
1190          */
1191         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1192                         f->file_name);
1193
1194         if (!fio_option_is_set(&td->o, random_generator)) {
1195                 log_info("fio: Switching to tausworthe64. Use the "
1196                          "random_generator= option to get rid of this "
1197                          "warning.\n");
1198                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1199                 return 0;
1200         }
1201
1202         /*
1203          * Just make this information to avoid breaking scripts.
1204          */
1205         log_info("fio: Use the random_generator= option to switch to lfsr or "
1206                          "tausworthe64.\n");
1207         return 0;
1208 }
1209
1210 int init_random_map(struct thread_data *td)
1211 {
1212         unsigned long long blocks;
1213         struct fio_file *f;
1214         unsigned int i;
1215
1216         if (init_rand_distribution(td))
1217                 return 0;
1218         if (!td_random(td))
1219                 return 0;
1220
1221         for_each_file(td, f, i) {
1222                 uint64_t fsize = min(f->real_file_size, f->io_size);
1223
1224                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1225
1226                 if (check_rand_gen_limits(td, f, blocks))
1227                         return 1;
1228
1229                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1230                         unsigned long seed;
1231
1232                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1233
1234                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1235                                 fio_file_set_lfsr(f);
1236                                 continue;
1237                         }
1238                 } else if (!td->o.norandommap) {
1239                         f->io_axmap = axmap_new(blocks);
1240                         if (f->io_axmap) {
1241                                 fio_file_set_axmap(f);
1242                                 continue;
1243                         }
1244                 } else if (td->o.norandommap)
1245                         continue;
1246
1247                 if (!td->o.softrandommap) {
1248                         log_err("fio: failed allocating random map. If running"
1249                                 " a large number of jobs, try the 'norandommap'"
1250                                 " option or set 'softrandommap'. Or give"
1251                                 " a larger --alloc-size to fio.\n");
1252                         return 1;
1253                 }
1254
1255                 log_info("fio: file %s failed allocating random map. Running "
1256                          "job without.\n", f->file_name);
1257         }
1258
1259         return 0;
1260 }
1261
1262 void close_files(struct thread_data *td)
1263 {
1264         struct fio_file *f;
1265         unsigned int i;
1266
1267         for_each_file(td, f, i) {
1268                 if (fio_file_open(f))
1269                         td_io_close_file(td, f);
1270         }
1271 }
1272
1273 void close_and_free_files(struct thread_data *td)
1274 {
1275         struct fio_file *f;
1276         unsigned int i;
1277
1278         dprint(FD_FILE, "close files\n");
1279
1280         for_each_file(td, f, i) {
1281                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1282                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1283                         td_io_unlink_file(td, f);
1284                 }
1285
1286                 if (fio_file_open(f))
1287                         td_io_close_file(td, f);
1288
1289                 remove_file_hash(f);
1290
1291                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1292                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1293                         td_io_unlink_file(td, f);
1294                 }
1295
1296                 sfree(f->file_name);
1297                 f->file_name = NULL;
1298                 if (fio_file_axmap(f)) {
1299                         axmap_free(f->io_axmap);
1300                         f->io_axmap = NULL;
1301                 }
1302                 sfree(f);
1303         }
1304
1305         td->o.filename = NULL;
1306         free(td->files);
1307         free(td->file_locks);
1308         td->files_index = 0;
1309         td->files = NULL;
1310         td->file_locks = NULL;
1311         td->o.file_lock_mode = FILE_LOCK_NONE;
1312         td->o.nr_files = 0;
1313 }
1314
1315 static void get_file_type(struct fio_file *f)
1316 {
1317         struct stat sb;
1318
1319         if (!strcmp(f->file_name, "-"))
1320                 f->filetype = FIO_TYPE_PIPE;
1321         else
1322                 f->filetype = FIO_TYPE_FILE;
1323
1324 #ifdef WIN32
1325         /* \\.\ is the device namespace in Windows, where every file is
1326          * a block device */
1327         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1328                 f->filetype = FIO_TYPE_BLOCK;
1329 #endif
1330
1331         if (!stat(f->file_name, &sb)) {
1332                 if (S_ISBLK(sb.st_mode))
1333                         f->filetype = FIO_TYPE_BLOCK;
1334                 else if (S_ISCHR(sb.st_mode))
1335                         f->filetype = FIO_TYPE_CHAR;
1336                 else if (S_ISFIFO(sb.st_mode))
1337                         f->filetype = FIO_TYPE_PIPE;
1338         }
1339 }
1340
1341 static bool __is_already_allocated(const char *fname, bool set)
1342 {
1343         struct flist_head *entry;
1344         bool ret;
1345
1346         ret = file_bloom_exists(fname, set);
1347         if (!ret)
1348                 return ret;
1349
1350         flist_for_each(entry, &filename_list) {
1351                 struct file_name *fn;
1352
1353                 fn = flist_entry(entry, struct file_name, list);
1354
1355                 if (!strcmp(fn->filename, fname))
1356                         return true;
1357         }
1358
1359         return false;
1360 }
1361
1362 static bool is_already_allocated(const char *fname)
1363 {
1364         bool ret;
1365
1366         fio_file_hash_lock();
1367         ret = __is_already_allocated(fname, false);
1368         fio_file_hash_unlock();
1369
1370         return ret;
1371 }
1372
1373 static void set_already_allocated(const char *fname)
1374 {
1375         struct file_name *fn;
1376
1377         fn = malloc(sizeof(struct file_name));
1378         fn->filename = strdup(fname);
1379
1380         fio_file_hash_lock();
1381         if (!__is_already_allocated(fname, true)) {
1382                 flist_add_tail(&fn->list, &filename_list);
1383                 fn = NULL;
1384         }
1385         fio_file_hash_unlock();
1386
1387         if (fn) {
1388                 free(fn->filename);
1389                 free(fn);
1390         }
1391 }
1392
1393 static void free_already_allocated(void)
1394 {
1395         struct flist_head *entry, *tmp;
1396         struct file_name *fn;
1397
1398         if (flist_empty(&filename_list))
1399                 return;
1400
1401         fio_file_hash_lock();
1402         flist_for_each_safe(entry, tmp, &filename_list) {
1403                 fn = flist_entry(entry, struct file_name, list);
1404                 free(fn->filename);
1405                 flist_del(&fn->list);
1406                 free(fn);
1407         }
1408
1409         fio_file_hash_unlock();
1410 }
1411
1412 static struct fio_file *alloc_new_file(struct thread_data *td)
1413 {
1414         struct fio_file *f;
1415
1416         f = smalloc(sizeof(*f));
1417         if (!f) {
1418                 assert(0);
1419                 return NULL;
1420         }
1421
1422         f->fd = -1;
1423         f->shadow_fd = -1;
1424         fio_file_reset(td, f);
1425         return f;
1426 }
1427
1428 bool exists_and_not_regfile(const char *filename)
1429 {
1430         struct stat sb;
1431
1432         if (lstat(filename, &sb) == -1)
1433                 return false;
1434
1435 #ifndef WIN32 /* NOT Windows */
1436         if (S_ISREG(sb.st_mode))
1437                 return false;
1438 #else
1439         /* \\.\ is the device namespace in Windows, where every file
1440          * is a device node */
1441         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1442                 return false;
1443 #endif
1444
1445         return true;
1446 }
1447
1448 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1449 {
1450         int cur_files = td->files_index;
1451         char file_name[PATH_MAX];
1452         struct fio_file *f;
1453         int len = 0;
1454
1455         dprint(FD_FILE, "add file %s\n", fname);
1456
1457         if (td->o.directory)
1458                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1459                                         td->o.unique_filename);
1460
1461         sprintf(file_name + len, "%s", fname);
1462
1463         /* clean cloned siblings using existing files */
1464         if (numjob && is_already_allocated(file_name) &&
1465             !exists_and_not_regfile(fname))
1466                 return 0;
1467
1468         f = alloc_new_file(td);
1469
1470         if (td->files_size <= td->files_index) {
1471                 unsigned int new_size = td->o.nr_files + 1;
1472
1473                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1474
1475                 td->files = realloc(td->files, new_size * sizeof(f));
1476                 if (td->files == NULL) {
1477                         log_err("fio: realloc OOM\n");
1478                         assert(0);
1479                 }
1480                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1481                         td->file_locks = realloc(td->file_locks, new_size);
1482                         if (!td->file_locks) {
1483                                 log_err("fio: realloc OOM\n");
1484                                 assert(0);
1485                         }
1486                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1487                 }
1488                 td->files_size = new_size;
1489         }
1490         td->files[cur_files] = f;
1491         f->fileno = cur_files;
1492
1493         /*
1494          * init function, io engine may not be loaded yet
1495          */
1496         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1497                 f->real_file_size = -1ULL;
1498
1499         f->file_name = smalloc_strdup(file_name);
1500         if (!f->file_name)
1501                 assert(0);
1502
1503         get_file_type(f);
1504
1505         switch (td->o.file_lock_mode) {
1506         case FILE_LOCK_NONE:
1507                 break;
1508         case FILE_LOCK_READWRITE:
1509                 f->rwlock = fio_rwlock_init();
1510                 break;
1511         case FILE_LOCK_EXCLUSIVE:
1512                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1513                 break;
1514         default:
1515                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1516                 assert(0);
1517         }
1518
1519         td->files_index++;
1520         if (f->filetype == FIO_TYPE_FILE)
1521                 td->nr_normal_files++;
1522
1523         set_already_allocated(file_name);
1524
1525         if (inc)
1526                 td->o.nr_files++;
1527
1528         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1529                                                         cur_files);
1530
1531         return cur_files;
1532 }
1533
1534 int add_file_exclusive(struct thread_data *td, const char *fname)
1535 {
1536         struct fio_file *f;
1537         unsigned int i;
1538
1539         for_each_file(td, f, i) {
1540                 if (!strcmp(f->file_name, fname))
1541                         return i;
1542         }
1543
1544         return add_file(td, fname, 0, 1);
1545 }
1546
1547 void get_file(struct fio_file *f)
1548 {
1549         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1550         assert(fio_file_open(f));
1551         f->references++;
1552 }
1553
1554 int put_file(struct thread_data *td, struct fio_file *f)
1555 {
1556         int f_ret = 0, ret = 0;
1557
1558         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1559
1560         if (!fio_file_open(f)) {
1561                 assert(f->fd == -1);
1562                 return 0;
1563         }
1564
1565         assert(f->references);
1566         if (--f->references)
1567                 return 0;
1568
1569         if (should_fsync(td) && td->o.fsync_on_close) {
1570                 f_ret = fsync(f->fd);
1571                 if (f_ret < 0)
1572                         f_ret = errno;
1573         }
1574
1575         if (td->io_ops->close_file)
1576                 ret = td->io_ops->close_file(td, f);
1577
1578         if (!ret)
1579                 ret = f_ret;
1580
1581         td->nr_open_files--;
1582         fio_file_clear_open(f);
1583         assert(f->fd == -1);
1584         return ret;
1585 }
1586
1587 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1588 {
1589         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1590                 return;
1591
1592         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1593                 if (ddir == DDIR_READ)
1594                         fio_rwlock_read(f->rwlock);
1595                 else
1596                         fio_rwlock_write(f->rwlock);
1597         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1598                 fio_mutex_down(f->lock);
1599
1600         td->file_locks[f->fileno] = td->o.file_lock_mode;
1601 }
1602
1603 void unlock_file(struct thread_data *td, struct fio_file *f)
1604 {
1605         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1606                 return;
1607
1608         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1609                 fio_rwlock_unlock(f->rwlock);
1610         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1611                 fio_mutex_up(f->lock);
1612
1613         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1614 }
1615
1616 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1617 {
1618         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1619                 return;
1620         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1621                 unlock_file(td, f);
1622 }
1623
1624 static int recurse_dir(struct thread_data *td, const char *dirname)
1625 {
1626         struct dirent *dir;
1627         int ret = 0;
1628         DIR *D;
1629
1630         D = opendir(dirname);
1631         if (!D) {
1632                 char buf[FIO_VERROR_SIZE];
1633
1634                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1635                 td_verror(td, errno, buf);
1636                 return 1;
1637         }
1638
1639         while ((dir = readdir(D)) != NULL) {
1640                 char full_path[PATH_MAX];
1641                 struct stat sb;
1642
1643                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1644                         continue;
1645
1646                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1647
1648                 if (lstat(full_path, &sb) == -1) {
1649                         if (errno != ENOENT) {
1650                                 td_verror(td, errno, "stat");
1651                                 ret = 1;
1652                                 break;
1653                         }
1654                 }
1655
1656                 if (S_ISREG(sb.st_mode)) {
1657                         add_file(td, full_path, 0, 1);
1658                         continue;
1659                 }
1660                 if (!S_ISDIR(sb.st_mode))
1661                         continue;
1662
1663                 ret = recurse_dir(td, full_path);
1664                 if (ret)
1665                         break;
1666         }
1667
1668         closedir(D);
1669         return ret;
1670 }
1671
1672 int add_dir_files(struct thread_data *td, const char *path)
1673 {
1674         int ret = recurse_dir(td, path);
1675
1676         if (!ret)
1677                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1678
1679         return ret;
1680 }
1681
1682 void dup_files(struct thread_data *td, struct thread_data *org)
1683 {
1684         struct fio_file *f;
1685         unsigned int i;
1686
1687         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1688
1689         if (!org->files)
1690                 return;
1691
1692         td->files = malloc(org->files_index * sizeof(f));
1693
1694         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1695                 td->file_locks = malloc(org->files_index);
1696
1697         for_each_file(org, f, i) {
1698                 struct fio_file *__f;
1699
1700                 __f = alloc_new_file(td);
1701
1702                 if (f->file_name) {
1703                         __f->file_name = smalloc_strdup(f->file_name);
1704                         if (!__f->file_name)
1705                                 assert(0);
1706
1707                         __f->filetype = f->filetype;
1708                 }
1709
1710                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1711                         __f->lock = f->lock;
1712                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1713                         __f->rwlock = f->rwlock;
1714
1715                 td->files[i] = __f;
1716         }
1717 }
1718
1719 /*
1720  * Returns the index that matches the filename, or -1 if not there
1721  */
1722 int get_fileno(struct thread_data *td, const char *fname)
1723 {
1724         struct fio_file *f;
1725         unsigned int i;
1726
1727         for_each_file(td, f, i)
1728                 if (!strcmp(f->file_name, fname))
1729                         return i;
1730
1731         return -1;
1732 }
1733
1734 /*
1735  * For log usage, where we add/open/close files automatically
1736  */
1737 void free_release_files(struct thread_data *td)
1738 {
1739         close_files(td);
1740         td->o.nr_files = 0;
1741         td->o.open_files = 0;
1742         td->files_index = 0;
1743         td->nr_normal_files = 0;
1744 }
1745
1746 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1747 {
1748         int i;
1749
1750         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1751                 f->last_pos[i] = f->file_offset;
1752                 f->last_start[i] = -1ULL;
1753         }
1754
1755         if (fio_file_axmap(f))
1756                 axmap_reset(f->io_axmap);
1757         else if (fio_file_lfsr(f))
1758                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1759 }
1760
1761 bool fio_files_done(struct thread_data *td)
1762 {
1763         struct fio_file *f;
1764         unsigned int i;
1765
1766         for_each_file(td, f, i)
1767                 if (!fio_file_done(f))
1768                         return false;
1769
1770         return true;
1771 }
1772
1773 /* free memory used in initialization phase only */
1774 void filesetup_mem_free(void)
1775 {
1776         free_already_allocated();
1777 }