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