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