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