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