filesetup.c

   1 #include <unistd.h>
   2 #include <fcntl.h>
   3 #include <string.h>
   4 #include <assert.h>
   5 #include <dirent.h>
   6 #include <libgen.h>
   7 #include <sys/stat.h>
   8 #include <sys/mman.h>
   9 #include <sys/types.h>
  10
  11 #include "fio.h"
  12 #include "smalloc.h"
  13 #include "filehash.h"
  14 #include "options.h"
  15 #include "os/os.h"
  16 #include "hash.h"
  17 #include "lib/axmap.h"
  18
  19 #ifdef CONFIG_LINUX_FALLOCATE
  20 #include <linux/falloc.h>
  21 #endif
  22
  23 static int root_warn;
  24
  25 static FLIST_HEAD(filename_list);
  26
  27 /*
  28  * List entry for filename_list
  29  */
  30 struct file_name {
  31         struct flist_head list;
  32         char *filename;
  33 };
  34
  35 static inline void clear_error(struct thread_data *td)
  36 {
  37         td->error = 0;
  38         td->verror[0] = '\0';
  39 }
  40
  41 /*
  42  * Leaves f->fd open on success, caller must close
  43  */
  44 static int extend_file(struct thread_data *td, struct fio_file *f)
  45 {
  46         int r, new_layout = 0, unlink_file = 0, flags;
  47         unsigned long long left;
  48         unsigned int bs;
  49         char *b = NULL;
  50
  51         if (read_only) {
  52                 log_err("fio: refusing extend of file due to read-only\n");
  53                 return 0;
  54         }
  55
  56         /*
  57          * check if we need to lay the file out complete again. fio
  58          * does that for operations involving reads, or for writes
  59          * where overwrite is set
  60          */
  61         if (td_read(td) ||
  62            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
  63             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
  64                 new_layout = 1;
  65         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
  66                 unlink_file = 1;
  67
  68         if (unlink_file || new_layout) {
  69                 int ret;
  70
  71                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  72
  73                 ret = td_io_unlink_file(td, f);
  74                 if (ret != 0 && ret != ENOENT) {
  75                         td_verror(td, errno, "unlink");
  76                         return 1;
  77                 }
  78         }
  79
  80         flags = O_WRONLY;
  81         if (td->o.allow_create)
  82                 flags |= O_CREAT;
  83         if (new_layout)
  84                 flags |= O_TRUNC;
  85
  86 #ifdef WIN32
  87         flags |= _O_BINARY;
  88 #endif
  89
  90         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  91         f->fd = open(f->file_name, flags, 0644);
  92         if (f->fd < 0) {
  93                 int err = errno;
  94
  95                 if (err == ENOENT && !td->o.allow_create)
  96                         log_err("fio: file creation disallowed by "
  97                                         "allow_file_create=0\n");
  98                 else
  99                         td_verror(td, err, "open");
 100                 return 1;
 101         }
 102
 103 #ifdef CONFIG_POSIX_FALLOCATE
 104         if (!td->o.fill_device) {
 105                 switch (td->o.fallocate_mode) {
 106                 case FIO_FALLOCATE_NONE:
 107                         break;
 108                 case FIO_FALLOCATE_POSIX:
 109                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
 110                                  f->file_name,
 111                                  (unsigned long long) f->real_file_size);
 112
 113                         r = posix_fallocate(f->fd, 0, f->real_file_size);
 114                         if (r > 0) {
 115                                 log_err("fio: posix_fallocate fails: %s\n",
 116                                                 strerror(r));
 117                         }
 118                         break;
 119 #ifdef CONFIG_LINUX_FALLOCATE
 120                 case FIO_FALLOCATE_KEEP_SIZE:
 121                         dprint(FD_FILE,
 122                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
 123                                 "file %s size %llu\n", f->file_name,
 124                                 (unsigned long long) f->real_file_size);
 125
 126                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
 127                                         f->real_file_size);
 128                         if (r != 0)
 129                                 td_verror(td, errno, "fallocate");
 130
 131                         break;
 132 #endif /* CONFIG_LINUX_FALLOCATE */
 133                 default:
 134                         log_err("fio: unknown fallocate mode: %d\n",
 135                                 td->o.fallocate_mode);
 136                         assert(0);
 137                 }
 138         }
 139 #endif /* CONFIG_POSIX_FALLOCATE */
 140
 141         /*
 142          * If our jobs don't require regular files initially, we're done.
 143          */
 144         if (!new_layout)
 145                 goto done;
 146
 147         /*
 148          * The size will be -1ULL when fill_device is used, so don't truncate
 149          * or fallocate this file, just write it
 150          */
 151         if (!td->o.fill_device) {
 152                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 153                                         (unsigned long long) f->real_file_size);
 154                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 155                         if (errno != EFBIG) {
 156                                 td_verror(td, errno, "ftruncate");
 157                                 goto err;
 158                         }
 159                 }
 160         }
 161
 162         b = malloc(td->o.max_bs[DDIR_WRITE]);
 163
 164         left = f->real_file_size;
 165         while (left && !td->terminate) {
 166                 bs = td->o.max_bs[DDIR_WRITE];
 167                 if (bs > left)
 168                         bs = left;
 169
 170                 fill_io_buffer(td, b, bs, bs);
 171
 172                 r = write(f->fd, b, bs);
 173
 174                 if (r > 0) {
 175                         left -= r;
 176                         continue;
 177                 } else {
 178                         if (r < 0) {
 179                                 int __e = errno;
 180
 181                                 if (__e == ENOSPC) {
 182                                         if (td->o.fill_device)
 183                                                 break;
 184                                         log_info("fio: ENOSPC on laying out "
 185                                                  "file, stopping\n");
 186                                         break;
 187                                 }
 188                                 td_verror(td, errno, "write");
 189                         } else
 190                                 td_verror(td, EIO, "write");
 191
 192                         break;
 193                 }
 194         }
 195
 196         if (td->terminate) {
 197                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 198                 td_io_unlink_file(td, f);
 199         } else if (td->o.create_fsync) {
 200                 if (fsync(f->fd) < 0) {
 201                         td_verror(td, errno, "fsync");
 202                         goto err;
 203                 }
 204         }
 205         if (td->o.fill_device && !td_write(td)) {
 206                 fio_file_clear_size_known(f);
 207                 if (td_io_get_file_size(td, f))
 208                         goto err;
 209                 if (f->io_size > f->real_file_size)
 210                         f->io_size = f->real_file_size;
 211         }
 212
 213         free(b);
 214 done:
 215         return 0;
 216 err:
 217         close(f->fd);
 218         f->fd = -1;
 219         if (b)
 220                 free(b);
 221         return 1;
 222 }
 223
 224 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 225 {
 226         int ret = 0, r, did_open = 0, old_runstate;
 227         unsigned long long left;
 228         unsigned int bs;
 229         char *b;
 230
 231         if (td_ioengine_flagged(td, FIO_PIPEIO))
 232                 return 0;
 233
 234         if (!fio_file_open(f)) {
 235                 if (td->io_ops->open_file(td, f)) {
 236                         log_err("fio: cannot pre-read, failed to open file\n");
 237                         return 1;
 238                 }
 239                 did_open = 1;
 240         }
 241
 242         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 243
 244         bs = td->o.max_bs[DDIR_READ];
 245         b = malloc(bs);
 246         memset(b, 0, bs);
 247
 248         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 249                 td_verror(td, errno, "lseek");
 250                 log_err("fio: failed to lseek pre-read file\n");
 251                 ret = 1;
 252                 goto error;
 253         }
 254
 255         left = f->io_size;
 256
 257         while (left && !td->terminate) {
 258                 if (bs > left)
 259                         bs = left;
 260
 261                 r = read(f->fd, b, bs);
 262
 263                 if (r == (int) bs) {
 264                         left -= bs;
 265                         continue;
 266                 } else {
 267                         td_verror(td, EIO, "pre_read");
 268                         break;
 269                 }
 270         }
 271
 272 error:
 273         td_restore_runstate(td, old_runstate);
 274
 275         if (did_open)
 276                 td->io_ops->close_file(td, f);
 277
 278         free(b);
 279         return ret;
 280 }
 281
 282 unsigned long long get_rand_file_size(struct thread_data *td)
 283 {
 284         unsigned long long ret, sized;
 285         uint64_t frand_max;
 286         unsigned long r;
 287
 288         frand_max = rand_max(&td->file_size_state);
 289         r = __rand(&td->file_size_state);
 290         sized = td->o.file_size_high - td->o.file_size_low;
 291         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 292         ret += td->o.file_size_low;
 293         ret -= (ret % td->o.rw_min_bs);
 294         return ret;
 295 }
 296
 297 static int file_size(struct thread_data *td, struct fio_file *f)
 298 {
 299         struct stat st;
 300
 301         if (stat(f->file_name, &st) == -1) {
 302                 td_verror(td, errno, "fstat");
 303                 return 1;
 304         }
 305
 306         f->real_file_size = st.st_size;
 307         return 0;
 308 }
 309
 310 static int bdev_size(struct thread_data *td, struct fio_file *f)
 311 {
 312         unsigned long long bytes = 0;
 313         int r;
 314
 315         if (td->io_ops->open_file(td, f)) {
 316                 log_err("fio: failed opening blockdev %s for size check\n",
 317                         f->file_name);
 318                 return 1;
 319         }
 320
 321         r = blockdev_size(f, &bytes);
 322         if (r) {
 323                 td_verror(td, r, "blockdev_size");
 324                 goto err;
 325         }
 326
 327         if (!bytes) {
 328                 log_err("%s: zero sized block device?\n", f->file_name);
 329                 goto err;
 330         }
 331
 332         f->real_file_size = bytes;
 333         td->io_ops->close_file(td, f);
 334         return 0;
 335 err:
 336         td->io_ops->close_file(td, f);
 337         return 1;
 338 }
 339
 340 static int char_size(struct thread_data *td, struct fio_file *f)
 341 {
 342 #ifdef FIO_HAVE_CHARDEV_SIZE
 343         unsigned long long bytes = 0;
 344         int r;
 345
 346         if (td->io_ops->open_file(td, f)) {
 347                 log_err("fio: failed opening chardev %s for size check\n",
 348                         f->file_name);
 349                 return 1;
 350         }
 351
 352         r = chardev_size(f, &bytes);
 353         if (r) {
 354                 td_verror(td, r, "chardev_size");
 355                 goto err;
 356         }
 357
 358         if (!bytes) {
 359                 log_err("%s: zero sized char device?\n", f->file_name);
 360                 goto err;
 361         }
 362
 363         f->real_file_size = bytes;
 364         td->io_ops->close_file(td, f);
 365         return 0;
 366 err:
 367         td->io_ops->close_file(td, f);
 368         return 1;
 369 #else
 370         f->real_file_size = -1ULL;
 371         return 0;
 372 #endif
 373 }
 374
 375 static int get_file_size(struct thread_data *td, struct fio_file *f)
 376 {
 377         int ret = 0;
 378
 379         if (fio_file_size_known(f))
 380                 return 0;
 381
 382         if (f->filetype == FIO_TYPE_FILE)
 383                 ret = file_size(td, f);
 384         else if (f->filetype == FIO_TYPE_BLOCK)
 385                 ret = bdev_size(td, f);
 386         else if (f->filetype == FIO_TYPE_CHAR)
 387                 ret = char_size(td, f);
 388         else
 389                 f->real_file_size = -1ULL;
 390
 391         /*
 392          * Leave ->real_file_size with 0 since it could be expectation
 393          * of initial setup for regular files.
 394          */
 395         if (ret)
 396                 return ret;
 397
 398         /*
 399          * If ->real_file_size is -1, a conditional for the message
 400          * "offset extends end" is always true, but it makes no sense,
 401          * so just return the same value here.
 402          */
 403         if (f->real_file_size == -1ULL) {
 404                 log_info("%s: failed to get file size of %s\n", td->o.name,
 405                                         f->file_name);
 406                 return 1;
 407         }
 408
 409         if (td->o.start_offset && f->file_offset == 0)
 410                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 411                                         f->file_name);
 412         /*
 413          * ->file_offset normally hasn't been initialized yet, so this
 414          * is basically always false.
 415          */
 416         if (f->file_offset > f->real_file_size) {
 417                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 418                                         (unsigned long long) f->file_offset,
 419                                         (unsigned long long) f->real_file_size);
 420                 return 1;
 421         }
 422
 423         fio_file_set_size_known(f);
 424         return 0;
 425 }
 426
 427 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 428                                    unsigned long long off,
 429                                    unsigned long long len)
 430 {
 431         int errval = 0, ret = 0;
 432
 433 #ifdef CONFIG_ESX
 434         return 0;
 435 #endif
 436
 437         if (len == -1ULL)
 438                 len = f->io_size;
 439         if (off == -1ULL)
 440                 off = f->file_offset;
 441
 442         if (len == -1ULL || off == -1ULL)
 443                 return 0;
 444
 445         if (td->io_ops->invalidate) {
 446                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 447                         f->file_name);
 448                 ret = td->io_ops->invalidate(td, f);
 449                 if (ret < 0)
 450                         errval = -ret;
 451         } else if (f->filetype == FIO_TYPE_FILE) {
 452                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 453                         f->file_name, off, len);
 454                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 455                 if (ret)
 456                         errval = ret;
 457         } else if (f->filetype == FIO_TYPE_BLOCK) {
 458                 int retry_count = 0;
 459
 460                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 461                 ret = blockdev_invalidate_cache(f);
 462                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 463                         /*
 464                          * Linux multipath devices reject ioctl while
 465                          * the maps are being updated. That window can
 466                          * last tens of milliseconds; we'll try up to
 467                          * a quarter of a second.
 468                          */
 469                         usleep(10000);
 470                         ret = blockdev_invalidate_cache(f);
 471                 }
 472                 if (ret < 0 && errno == EACCES && geteuid()) {
 473                         if (!root_warn) {
 474                                 log_err("fio: only root may flush block "
 475                                         "devices. Cache flush bypassed!\n");
 476                                 root_warn = 1;
 477                         }
 478                         ret = 0;
 479                 }
 480                 if (ret < 0)
 481                         errval = errno;
 482                 else if (ret) /* probably not supported */
 483                         errval = ret;
 484         } else if (f->filetype == FIO_TYPE_CHAR ||
 485                    f->filetype == FIO_TYPE_PIPE) {
 486                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 487                 ret = 0;
 488         }
 489
 490         /*
 491          * Cache flushing isn't a fatal condition, and we know it will
 492          * happen on some platforms where we don't have the proper
 493          * function to flush eg block device caches. So just warn and
 494          * continue on our way.
 495          */
 496         if (errval)
 497                 log_info("fio: cache invalidation of %s failed: %s\n",
 498                          f->file_name, strerror(errval));
 499
 500         return 0;
 501
 502 }
 503
 504 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 505 {
 506         if (!fio_file_open(f))
 507                 return 0;
 508
 509         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 510 }
 511
 512 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 513 {
 514         int ret = 0;
 515
 516         dprint(FD_FILE, "fd close %s\n", f->file_name);
 517
 518         remove_file_hash(f);
 519
 520         if (close(f->fd) < 0)
 521                 ret = errno;
 522
 523         f->fd = -1;
 524
 525         if (f->shadow_fd != -1) {
 526                 close(f->shadow_fd);
 527                 f->shadow_fd = -1;
 528         }
 529
 530         f->engine_pos = 0;
 531         return ret;
 532 }
 533
 534 int file_lookup_open(struct fio_file *f, int flags)
 535 {
 536         struct fio_file *__f;
 537         int from_hash;
 538
 539         __f = lookup_file_hash(f->file_name);
 540         if (__f) {
 541                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 542                 f->lock = __f->lock;
 543                 from_hash = 1;
 544         } else {
 545                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 546                 from_hash = 0;
 547         }
 548
 549 #ifdef WIN32
 550         flags |= _O_BINARY;
 551 #endif
 552
 553         f->fd = open(f->file_name, flags, 0600);
 554         return from_hash;
 555 }
 556
 557 static int file_close_shadow_fds(struct thread_data *td)
 558 {
 559         struct fio_file *f;
 560         int num_closed = 0;
 561         unsigned int i;
 562
 563         for_each_file(td, f, i) {
 564                 if (f->shadow_fd == -1)
 565                         continue;
 566
 567                 close(f->shadow_fd);
 568                 f->shadow_fd = -1;
 569                 num_closed++;
 570         }
 571
 572         return num_closed;
 573 }
 574
 575 int generic_open_file(struct thread_data *td, struct fio_file *f)
 576 {
 577         int is_std = 0;
 578         int flags = 0;
 579         int from_hash = 0;
 580
 581         dprint(FD_FILE, "fd open %s\n", f->file_name);
 582
 583         if (!strcmp(f->file_name, "-")) {
 584                 if (td_rw(td)) {
 585                         log_err("fio: can't read/write to stdin/out\n");
 586                         return 1;
 587                 }
 588                 is_std = 1;
 589
 590                 /*
 591                  * move output logging to stderr, if we are writing to stdout
 592                  */
 593                 if (td_write(td))
 594                         f_out = stderr;
 595         }
 596
 597         if (td_trim(td))
 598                 goto skip_flags;
 599         if (td->o.odirect)
 600                 flags |= OS_O_DIRECT;
 601         if (td->o.oatomic) {
 602                 if (!FIO_O_ATOMIC) {
 603                         td_verror(td, EINVAL, "OS does not support atomic IO");
 604                         return 1;
 605                 }
 606                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 607         }
 608         if (td->o.sync_io)
 609                 flags |= O_SYNC;
 610         if (td->o.create_on_open && td->o.allow_create)
 611                 flags |= O_CREAT;
 612 skip_flags:
 613         if (f->filetype != FIO_TYPE_FILE)
 614                 flags |= FIO_O_NOATIME;
 615
 616 open_again:
 617         if (td_write(td)) {
 618                 if (!read_only)
 619                         flags |= O_RDWR;
 620
 621                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 622                         flags |= O_CREAT;
 623
 624                 if (is_std)
 625                         f->fd = dup(STDOUT_FILENO);
 626                 else
 627                         from_hash = file_lookup_open(f, flags);
 628         } else if (td_read(td)) {
 629                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 630                         flags |= O_RDWR;
 631                 else
 632                         flags |= O_RDONLY;
 633
 634                 if (is_std)
 635                         f->fd = dup(STDIN_FILENO);
 636                 else
 637                         from_hash = file_lookup_open(f, flags);
 638         } else if (td_trim(td)) {
 639                 assert(!td_rw(td)); /* should have matched above */
 640                 flags |= O_RDWR;
 641                 from_hash = file_lookup_open(f, flags);
 642         }
 643
 644         if (f->fd == -1) {
 645                 char buf[FIO_VERROR_SIZE];
 646                 int __e = errno;
 647
 648                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 649                         flags &= ~FIO_O_NOATIME;
 650                         goto open_again;
 651                 }
 652                 if (__e == EMFILE && file_close_shadow_fds(td))
 653                         goto open_again;
 654
 655                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 656
 657                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 658                         log_err("fio: looks like your file system does not " \
 659                                 "support direct=1/buffered=0\n");
 660                 }
 661
 662                 td_verror(td, __e, buf);
 663                 return 1;
 664         }
 665
 666         if (!from_hash && f->fd != -1) {
 667                 if (add_file_hash(f)) {
 668                         int fio_unused ret;
 669
 670                         /*
 671                          * Stash away descriptor for later close. This is to
 672                          * work-around a "feature" on Linux, where a close of
 673                          * an fd that has been opened for write will trigger
 674                          * udev to call blkid to check partitions, fs id, etc.
 675                          * That pollutes the device cache, which can slow down
 676                          * unbuffered accesses.
 677                          */
 678                         if (f->shadow_fd == -1)
 679                                 f->shadow_fd = f->fd;
 680                         else {
 681                                 /*
 682                                  * OK to ignore, we haven't done anything
 683                                  * with it
 684                                  */
 685                                 ret = generic_close_file(td, f);
 686                         }
 687                         goto open_again;
 688                 }
 689         }
 690
 691         return 0;
 692 }
 693
 694 /*
 695  * This function i.e. get_file_size() is the default .get_file_size
 696  * implementation of majority of I/O engines.
 697  */
 698 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 699 {
 700         return get_file_size(td, f);
 701 }
 702
 703 /*
 704  * open/close all files, so that ->real_file_size gets set
 705  */
 706 static int get_file_sizes(struct thread_data *td)
 707 {
 708         struct fio_file *f;
 709         unsigned int i;
 710         int err = 0;
 711
 712         for_each_file(td, f, i) {
 713                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 714                                                                 f->file_name);
 715
 716                 if (td_io_get_file_size(td, f)) {
 717                         if (td->error != ENOENT) {
 718                                 log_err("%s\n", td->verror);
 719                                 err = 1;
 720                                 break;
 721                         }
 722                         clear_error(td);
 723                 }
 724
 725                 /*
 726                  * There are corner cases where we end up with -1 for
 727                  * ->real_file_size due to unsupported file type, etc.
 728                  * We then just set to size option value divided by number
 729                  * of files, similar to the way file ->io_size is set.
 730                  * stat(2) failure doesn't set ->real_file_size to -1.
 731                  */
 732                 if (f->real_file_size == -1ULL && td->o.size)
 733                         f->real_file_size = td->o.size / td->o.nr_files;
 734         }
 735
 736         return err;
 737 }
 738
 739 struct fio_mount {
 740         struct flist_head list;
 741         const char *base;
 742         char __base[256];
 743         unsigned int key;
 744 };
 745
 746 /*
 747  * Get free number of bytes for each file on each unique mount.
 748  */
 749 static unsigned long long get_fs_free_counts(struct thread_data *td)
 750 {
 751         struct flist_head *n, *tmp;
 752         unsigned long long ret = 0;
 753         struct fio_mount *fm;
 754         FLIST_HEAD(list);
 755         struct fio_file *f;
 756         unsigned int i;
 757
 758         for_each_file(td, f, i) {
 759                 struct stat sb;
 760                 char buf[256];
 761
 762                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 763                         if (f->real_file_size != -1ULL)
 764                                 ret += f->real_file_size;
 765                         continue;
 766                 } else if (f->filetype != FIO_TYPE_FILE)
 767                         continue;
 768
 769                 buf[255] = '\0';
 770                 strncpy(buf, f->file_name, 255);
 771
 772                 if (stat(buf, &sb) < 0) {
 773                         if (errno != ENOENT)
 774                                 break;
 775                         strcpy(buf, ".");
 776                         if (stat(buf, &sb) < 0)
 777                                 break;
 778                 }
 779
 780                 fm = NULL;
 781                 flist_for_each(n, &list) {
 782                         fm = flist_entry(n, struct fio_mount, list);
 783                         if (fm->key == sb.st_dev)
 784                                 break;
 785
 786                         fm = NULL;
 787                 }
 788
 789                 if (fm)
 790                         continue;
 791
 792                 fm = calloc(1, sizeof(*fm));
 793                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 794                 fm->base = basename(fm->__base);
 795                 fm->key = sb.st_dev;
 796                 flist_add(&fm->list, &list);
 797         }
 798
 799         flist_for_each_safe(n, tmp, &list) {
 800                 unsigned long long sz;
 801
 802                 fm = flist_entry(n, struct fio_mount, list);
 803                 flist_del(&fm->list);
 804
 805                 sz = get_fs_free_size(fm->base);
 806                 if (sz && sz != -1ULL)
 807                         ret += sz;
 808
 809                 free(fm);
 810         }
 811
 812         return ret;
 813 }
 814
 815 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 816 {
 817         struct thread_options *o = &td->o;
 818
 819         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 820                 return f->real_file_size;
 821
 822         return td->o.start_offset +
 823                 td->subjob_number * td->o.offset_increment;
 824 }
 825
 826 /*
 827  * Open the files and setup files sizes, creating files if necessary.
 828  */
 829 int setup_files(struct thread_data *td)
 830 {
 831         unsigned long long total_size, extend_size;
 832         struct thread_options *o = &td->o;
 833         struct fio_file *f;
 834         unsigned int i, nr_fs_extra = 0;
 835         int err = 0, need_extend;
 836         int old_state;
 837         const unsigned int bs = td_min_bs(td);
 838         uint64_t fs = 0;
 839
 840         dprint(FD_FILE, "setup files\n");
 841
 842         old_state = td_bump_runstate(td, TD_SETTING_UP);
 843
 844         if (o->read_iolog_file)
 845                 goto done;
 846
 847         /*
 848          * Find out physical size of files or devices for this thread,
 849          * before we determine I/O size and range of our targets.
 850          * If ioengine defines a setup() method, it's responsible for
 851          * opening the files and setting f->real_file_size to indicate
 852          * the valid range for that file.
 853          */
 854         if (td->io_ops->setup)
 855                 err = td->io_ops->setup(td);
 856         else
 857                 err = get_file_sizes(td);
 858
 859         if (err)
 860                 goto err_out;
 861
 862         /*
 863          * check sizes. if the files/devices do not exist and the size
 864          * isn't passed to fio, abort.
 865          */
 866         total_size = 0;
 867         for_each_file(td, f, i) {
 868                 f->fileno = i;
 869                 if (f->real_file_size == -1ULL)
 870                         total_size = -1ULL;
 871                 else
 872                         total_size += f->real_file_size;
 873         }
 874
 875         if (o->fill_device)
 876                 td->fill_device_size = get_fs_free_counts(td);
 877
 878         /*
 879          * device/file sizes are zero and no size given, punt
 880          */
 881         if ((!total_size || total_size == -1ULL) && !o->size &&
 882             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 883             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 884                 log_err("%s: you need to specify size=\n", o->name);
 885                 td_verror(td, EINVAL, "total_file_size");
 886                 goto err_out;
 887         }
 888
 889         /*
 890          * Calculate per-file size and potential extra size for the
 891          * first files, if needed (i.e. if we don't have a fixed size).
 892          */
 893         if (!o->file_size_low && o->nr_files) {
 894                 uint64_t all_fs;
 895
 896                 fs = o->size / o->nr_files;
 897                 all_fs = fs * o->nr_files;
 898
 899                 if (all_fs < o->size)
 900                         nr_fs_extra = (o->size - all_fs) / bs;
 901         }
 902
 903         /*
 904          * now file sizes are known, so we can set ->io_size. if size= is
 905          * not given, ->io_size is just equal to ->real_file_size. if size
 906          * is given, ->io_size is size / nr_files.
 907          */
 908         extend_size = total_size = 0;
 909         need_extend = 0;
 910         for_each_file(td, f, i) {
 911                 f->file_offset = get_start_offset(td, f);
 912
 913                 /*
 914                  * Update ->io_size depending on options specified.
 915                  * ->file_size_low being 0 means filesize option isn't set.
 916                  * Non zero ->file_size_low equals ->file_size_high means
 917                  * filesize option is set in a fixed size format.
 918                  * Non zero ->file_size_low not equals ->file_size_high means
 919                  * filesize option is set in a range format.
 920                  */
 921                 if (!o->file_size_low) {
 922                         /*
 923                          * no file size or range given, file size is equal to
 924                          * total size divided by number of files. If the size
 925                          * doesn't divide nicely with the min blocksize,
 926                          * make the first files bigger.
 927                          */
 928                         f->io_size = fs;
 929                         if (nr_fs_extra) {
 930                                 nr_fs_extra--;
 931                                 f->io_size += bs;
 932                         }
 933
 934                         /*
 935                          * We normally don't come here, but if the result is 0,
 936                          * set it to the real file size. This could be size of
 937                          * the existing one if it already exists, but otherwise
 938                          * will be set to 0. A new file won't be created because
 939                          * ->io_size + ->file_offset equals ->real_file_size.
 940                          */
 941                         if (!f->io_size) {
 942                                 if (f->file_offset > f->real_file_size)
 943                                         goto err_offset;
 944                                 f->io_size = f->real_file_size - f->file_offset;
 945                                 log_info("fio: forcing file %s size to %llu\n",
 946                                         f->file_name,
 947                                         (unsigned long long)f->io_size);
 948                                 if (!f->io_size)
 949                                         log_info("fio: file %s may be ignored\n",
 950                                                 f->file_name);
 951                         }
 952                 } else if (f->real_file_size < o->file_size_low ||
 953                            f->real_file_size > o->file_size_high) {
 954                         if (f->file_offset > o->file_size_low)
 955                                 goto err_offset;
 956                         /*
 957                          * file size given. if it's fixed, use that. if it's a
 958                          * range, generate a random size in-between.
 959                          */
 960                         if (o->file_size_low == o->file_size_high)
 961                                 f->io_size = o->file_size_low - f->file_offset;
 962                         else {
 963                                 f->io_size = get_rand_file_size(td)
 964                                                 - f->file_offset;
 965                         }
 966                 } else
 967                         f->io_size = f->real_file_size - f->file_offset;
 968
 969                 if (f->io_size == -1ULL)
 970                         total_size = -1ULL;
 971                 else {
 972                         if (o->size_percent) {
 973                                 f->io_size = (f->io_size * o->size_percent) / 100;
 974                                 f->io_size -= (f->io_size % td_min_bs(td));
 975                         }
 976                         total_size += f->io_size;
 977                 }
 978
 979                 if (f->filetype == FIO_TYPE_FILE &&
 980                     (f->io_size + f->file_offset) > f->real_file_size &&
 981                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 982                         if (!o->create_on_open) {
 983                                 need_extend++;
 984                                 extend_size += (f->io_size + f->file_offset);
 985                                 fio_file_set_extend(f);
 986                         } else
 987                                 f->real_file_size = f->io_size + f->file_offset;
 988                 }
 989         }
 990
 991         if (td->o.block_error_hist) {
 992                 int len;
 993
 994                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
 995                 f = td->files[0];
 996                 len = f->io_size / td->o.bs[DDIR_TRIM];
 997                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
 998                         log_err("fio: cannot calculate block histogram with "
 999                                 "%d trim blocks, maximum %d\n",
1000                                 len, MAX_NR_BLOCK_INFOS);
1001                         td_verror(td, EINVAL, "block_error_hist");
1002                         goto err_out;
1003                 }
1004
1005                 td->ts.nr_block_infos = len;
1006                 for (i = 0; i < len; i++)
1007                         td->ts.block_infos[i] =
1008                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1009         } else
1010                 td->ts.nr_block_infos = 0;
1011
1012         if (!o->size || (total_size && o->size > total_size))
1013                 o->size = total_size;
1014
1015         if (o->size < td_min_bs(td)) {
1016                 log_err("fio: blocksize too large for data set\n");
1017                 goto err_out;
1018         }
1019
1020         /*
1021          * See if we need to extend some files, typically needed when our
1022          * target regular files don't exist yet, but our jobs require them
1023          * initially due to read I/Os.
1024          */
1025         if (need_extend) {
1026                 temp_stall_ts = 1;
1027                 if (output_format & FIO_OUTPUT_NORMAL) {
1028                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1029                                  o->name,
1030                                  need_extend > 1 ? "s" : "",
1031                                  need_extend,
1032                                  need_extend > 1 ? "s" : "",
1033                                  need_extend > 1 ? "total " : "",
1034                                  extend_size >> 20);
1035                 }
1036
1037                 for_each_file(td, f, i) {
1038                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1039
1040                         if (!fio_file_extend(f))
1041                                 continue;
1042
1043                         assert(f->filetype == FIO_TYPE_FILE);
1044                         fio_file_clear_extend(f);
1045                         if (!o->fill_device) {
1046                                 old_len = f->real_file_size;
1047                                 extend_len = f->io_size + f->file_offset -
1048                                                 old_len;
1049                         }
1050                         f->real_file_size = (f->io_size + f->file_offset);
1051                         err = extend_file(td, f);
1052                         if (err)
1053                                 break;
1054
1055                         err = __file_invalidate_cache(td, f, old_len,
1056                                                                 extend_len);
1057
1058                         /*
1059                          * Shut up static checker
1060                          */
1061                         if (f->fd != -1)
1062                                 close(f->fd);
1063
1064                         f->fd = -1;
1065                         if (err)
1066                                 break;
1067                 }
1068                 temp_stall_ts = 0;
1069         }
1070
1071         if (err)
1072                 goto err_out;
1073
1074         if (!o->zone_size)
1075                 o->zone_size = o->size;
1076
1077         /*
1078          * iolog already set the total io size, if we read back
1079          * stored entries.
1080          */
1081         if (!o->read_iolog_file) {
1082                 if (o->io_size)
1083                         td->total_io_size = o->io_size * o->loops;
1084                 else
1085                         td->total_io_size = o->size * o->loops;
1086         }
1087
1088 done:
1089         if (o->create_only)
1090                 td->done = 1;
1091
1092         td_restore_runstate(td, old_state);
1093         return 0;
1094 err_offset:
1095         log_err("%s: you need to specify valid offset=\n", o->name);
1096 err_out:
1097         td_restore_runstate(td, old_state);
1098         return 1;
1099 }
1100
1101 int pre_read_files(struct thread_data *td)
1102 {
1103         struct fio_file *f;
1104         unsigned int i;
1105
1106         dprint(FD_FILE, "pre_read files\n");
1107
1108         for_each_file(td, f, i) {
1109                 pre_read_file(td, f);
1110         }
1111
1112         return 1;
1113 }
1114
1115 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1116 {
1117         unsigned int range_size, seed;
1118         unsigned long nranges;
1119         uint64_t fsize;
1120
1121         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1122         fsize = min(f->real_file_size, f->io_size);
1123
1124         nranges = (fsize + range_size - 1) / range_size;
1125
1126         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1127         if (!td->o.rand_repeatable)
1128                 seed = td->rand_seeds[4];
1129
1130         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1131                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1132         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1133                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1134         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1135                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1136
1137         return 1;
1138 }
1139
1140 static int init_rand_distribution(struct thread_data *td)
1141 {
1142         struct fio_file *f;
1143         unsigned int i;
1144         int state;
1145
1146         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1147                 return 0;
1148
1149         state = td_bump_runstate(td, TD_SETTING_UP);
1150
1151         for_each_file(td, f, i)
1152                 __init_rand_distribution(td, f);
1153
1154         td_restore_runstate(td, state);
1155
1156         return 1;
1157 }
1158
1159 /*
1160  * Check if the number of blocks exceeds the randomness capability of
1161  * the selected generator. Tausworthe is 32-bit, the others are fullly
1162  * 64-bit capable.
1163  */
1164 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1165                                  uint64_t blocks)
1166 {
1167         if (blocks <= FRAND32_MAX)
1168                 return 0;
1169         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1170                 return 0;
1171
1172         /*
1173          * If the user hasn't specified a random generator, switch
1174          * to tausworthe64 with informational warning. If the user did
1175          * specify one, just warn.
1176          */
1177         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1178                         f->file_name);
1179
1180         if (!fio_option_is_set(&td->o, random_generator)) {
1181                 log_info("fio: Switching to tausworthe64. Use the "
1182                          "random_generator= option to get rid of this "
1183                          "warning.\n");
1184                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1185                 return 0;
1186         }
1187
1188         /*
1189          * Just make this information to avoid breaking scripts.
1190          */
1191         log_info("fio: Use the random_generator= option to switch to lfsr or "
1192                          "tausworthe64.\n");
1193         return 0;
1194 }
1195
1196 int init_random_map(struct thread_data *td)
1197 {
1198         unsigned long long blocks;
1199         struct fio_file *f;
1200         unsigned int i;
1201
1202         if (init_rand_distribution(td))
1203                 return 0;
1204         if (!td_random(td))
1205                 return 0;
1206
1207         for_each_file(td, f, i) {
1208                 uint64_t fsize = min(f->real_file_size, f->io_size);
1209
1210                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1211
1212                 if (check_rand_gen_limits(td, f, blocks))
1213                         return 1;
1214
1215                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1216                         unsigned long seed;
1217
1218                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1219
1220                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1221                                 fio_file_set_lfsr(f);
1222                                 continue;
1223                         }
1224                 } else if (!td->o.norandommap) {
1225                         f->io_axmap = axmap_new(blocks);
1226                         if (f->io_axmap) {
1227                                 fio_file_set_axmap(f);
1228                                 continue;
1229                         }
1230                 } else if (td->o.norandommap)
1231                         continue;
1232
1233                 if (!td->o.softrandommap) {
1234                         log_err("fio: failed allocating random map. If running"
1235                                 " a large number of jobs, try the 'norandommap'"
1236                                 " option or set 'softrandommap'. Or give"
1237                                 " a larger --alloc-size to fio.\n");
1238                         return 1;
1239                 }
1240
1241                 log_info("fio: file %s failed allocating random map. Running "
1242                          "job without.\n", f->file_name);
1243         }
1244
1245         return 0;
1246 }
1247
1248 void close_files(struct thread_data *td)
1249 {
1250         struct fio_file *f;
1251         unsigned int i;
1252
1253         for_each_file(td, f, i) {
1254                 if (fio_file_open(f))
1255                         td_io_close_file(td, f);
1256         }
1257 }
1258
1259 void close_and_free_files(struct thread_data *td)
1260 {
1261         struct fio_file *f;
1262         unsigned int i;
1263
1264         dprint(FD_FILE, "close files\n");
1265
1266         for_each_file(td, f, i) {
1267                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1268                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1269                         td_io_unlink_file(td, f);
1270                 }
1271
1272                 if (fio_file_open(f))
1273                         td_io_close_file(td, f);
1274
1275                 remove_file_hash(f);
1276
1277                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1278                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1279                         td_io_unlink_file(td, f);
1280                 }
1281
1282                 sfree(f->file_name);
1283                 f->file_name = NULL;
1284                 if (fio_file_axmap(f)) {
1285                         axmap_free(f->io_axmap);
1286                         f->io_axmap = NULL;
1287                 }
1288                 sfree(f);
1289         }
1290
1291         td->o.filename = NULL;
1292         free(td->files);
1293         free(td->file_locks);
1294         td->files_index = 0;
1295         td->files = NULL;
1296         td->file_locks = NULL;
1297         td->o.file_lock_mode = FILE_LOCK_NONE;
1298         td->o.nr_files = 0;
1299 }
1300
1301 static void get_file_type(struct fio_file *f)
1302 {
1303         struct stat sb;
1304
1305         if (!strcmp(f->file_name, "-"))
1306                 f->filetype = FIO_TYPE_PIPE;
1307         else
1308                 f->filetype = FIO_TYPE_FILE;
1309
1310 #ifdef WIN32
1311         /* \\.\ is the device namespace in Windows, where every file is
1312          * a block device */
1313         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1314                 f->filetype = FIO_TYPE_BLOCK;
1315 #endif
1316
1317         if (!stat(f->file_name, &sb)) {
1318                 if (S_ISBLK(sb.st_mode))
1319                         f->filetype = FIO_TYPE_BLOCK;
1320                 else if (S_ISCHR(sb.st_mode))
1321                         f->filetype = FIO_TYPE_CHAR;
1322                 else if (S_ISFIFO(sb.st_mode))
1323                         f->filetype = FIO_TYPE_PIPE;
1324         }
1325 }
1326
1327 static bool __is_already_allocated(const char *fname, bool set)
1328 {
1329         struct flist_head *entry;
1330         bool ret;
1331
1332         ret = file_bloom_exists(fname, set);
1333         if (!ret)
1334                 return ret;
1335
1336         flist_for_each(entry, &filename_list) {
1337                 struct file_name *fn;
1338
1339                 fn = flist_entry(entry, struct file_name, list);
1340
1341                 if (!strcmp(fn->filename, fname))
1342                         return true;
1343         }
1344
1345         return false;
1346 }
1347
1348 static bool is_already_allocated(const char *fname)
1349 {
1350         bool ret;
1351
1352         fio_file_hash_lock();
1353         ret = __is_already_allocated(fname, false);
1354         fio_file_hash_unlock();
1355
1356         return ret;
1357 }
1358
1359 static void set_already_allocated(const char *fname)
1360 {
1361         struct file_name *fn;
1362
1363         fn = malloc(sizeof(struct file_name));
1364         fn->filename = strdup(fname);
1365
1366         fio_file_hash_lock();
1367         if (!__is_already_allocated(fname, true)) {
1368                 flist_add_tail(&fn->list, &filename_list);
1369                 fn = NULL;
1370         }
1371         fio_file_hash_unlock();
1372
1373         if (fn) {
1374                 free(fn->filename);
1375                 free(fn);
1376         }
1377 }
1378
1379 static void free_already_allocated(void)
1380 {
1381         struct flist_head *entry, *tmp;
1382         struct file_name *fn;
1383
1384         if (flist_empty(&filename_list))
1385                 return;
1386
1387         fio_file_hash_lock();
1388         flist_for_each_safe(entry, tmp, &filename_list) {
1389                 fn = flist_entry(entry, struct file_name, list);
1390                 free(fn->filename);
1391                 flist_del(&fn->list);
1392                 free(fn);
1393         }
1394
1395         fio_file_hash_unlock();
1396 }
1397
1398 static struct fio_file *alloc_new_file(struct thread_data *td)
1399 {
1400         struct fio_file *f;
1401
1402         f = smalloc(sizeof(*f));
1403         if (!f) {
1404                 assert(0);
1405                 return NULL;
1406         }
1407
1408         f->fd = -1;
1409         f->shadow_fd = -1;
1410         fio_file_reset(td, f);
1411         return f;
1412 }
1413
1414 bool exists_and_not_regfile(const char *filename)
1415 {
1416         struct stat sb;
1417
1418         if (lstat(filename, &sb) == -1)
1419                 return false;
1420
1421 #ifndef WIN32 /* NOT Windows */
1422         if (S_ISREG(sb.st_mode))
1423                 return false;
1424 #else
1425         /* \\.\ is the device namespace in Windows, where every file
1426          * is a device node */
1427         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1428                 return false;
1429 #endif
1430
1431         return true;
1432 }
1433
1434 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1435 {
1436         int cur_files = td->files_index;
1437         char file_name[PATH_MAX];
1438         struct fio_file *f;
1439         int len = 0;
1440
1441         dprint(FD_FILE, "add file %s\n", fname);
1442
1443         if (td->o.directory)
1444                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1445                                         td->o.unique_filename);
1446
1447         sprintf(file_name + len, "%s", fname);
1448
1449         /* clean cloned siblings using existing files */
1450         if (numjob && is_already_allocated(file_name) &&
1451             !exists_and_not_regfile(fname))
1452                 return 0;
1453
1454         f = alloc_new_file(td);
1455
1456         if (td->files_size <= td->files_index) {
1457                 unsigned int new_size = td->o.nr_files + 1;
1458
1459                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1460
1461                 td->files = realloc(td->files, new_size * sizeof(f));
1462                 if (td->files == NULL) {
1463                         log_err("fio: realloc OOM\n");
1464                         assert(0);
1465                 }
1466                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1467                         td->file_locks = realloc(td->file_locks, new_size);
1468                         if (!td->file_locks) {
1469                                 log_err("fio: realloc OOM\n");
1470                                 assert(0);
1471                         }
1472                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1473                 }
1474                 td->files_size = new_size;
1475         }
1476         td->files[cur_files] = f;
1477         f->fileno = cur_files;
1478
1479         /*
1480          * init function, io engine may not be loaded yet
1481          */
1482         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1483                 f->real_file_size = -1ULL;
1484
1485         f->file_name = smalloc_strdup(file_name);
1486         if (!f->file_name)
1487                 assert(0);
1488
1489         get_file_type(f);
1490
1491         switch (td->o.file_lock_mode) {
1492         case FILE_LOCK_NONE:
1493                 break;
1494         case FILE_LOCK_READWRITE:
1495                 f->rwlock = fio_rwlock_init();
1496                 break;
1497         case FILE_LOCK_EXCLUSIVE:
1498                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1499                 break;
1500         default:
1501                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1502                 assert(0);
1503         }
1504
1505         td->files_index++;
1506         if (f->filetype == FIO_TYPE_FILE)
1507                 td->nr_normal_files++;
1508
1509         set_already_allocated(file_name);
1510
1511         if (inc)
1512                 td->o.nr_files++;
1513
1514         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1515                                                         cur_files);
1516
1517         return cur_files;
1518 }
1519
1520 int add_file_exclusive(struct thread_data *td, const char *fname)
1521 {
1522         struct fio_file *f;
1523         unsigned int i;
1524
1525         for_each_file(td, f, i) {
1526                 if (!strcmp(f->file_name, fname))
1527                         return i;
1528         }
1529
1530         return add_file(td, fname, 0, 1);
1531 }
1532
1533 void get_file(struct fio_file *f)
1534 {
1535         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1536         assert(fio_file_open(f));
1537         f->references++;
1538 }
1539
1540 int put_file(struct thread_data *td, struct fio_file *f)
1541 {
1542         int f_ret = 0, ret = 0;
1543
1544         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1545
1546         if (!fio_file_open(f)) {
1547                 assert(f->fd == -1);
1548                 return 0;
1549         }
1550
1551         assert(f->references);
1552         if (--f->references)
1553                 return 0;
1554
1555         if (should_fsync(td) && td->o.fsync_on_close) {
1556                 f_ret = fsync(f->fd);
1557                 if (f_ret < 0)
1558                         f_ret = errno;
1559         }
1560
1561         if (td->io_ops->close_file)
1562                 ret = td->io_ops->close_file(td, f);
1563
1564         if (!ret)
1565                 ret = f_ret;
1566
1567         td->nr_open_files--;
1568         fio_file_clear_open(f);
1569         assert(f->fd == -1);
1570         return ret;
1571 }
1572
1573 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1574 {
1575         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1576                 return;
1577
1578         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1579                 if (ddir == DDIR_READ)
1580                         fio_rwlock_read(f->rwlock);
1581                 else
1582                         fio_rwlock_write(f->rwlock);
1583         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1584                 fio_mutex_down(f->lock);
1585
1586         td->file_locks[f->fileno] = td->o.file_lock_mode;
1587 }
1588
1589 void unlock_file(struct thread_data *td, struct fio_file *f)
1590 {
1591         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1592                 return;
1593
1594         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1595                 fio_rwlock_unlock(f->rwlock);
1596         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1597                 fio_mutex_up(f->lock);
1598
1599         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1600 }
1601
1602 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1603 {
1604         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1605                 return;
1606         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1607                 unlock_file(td, f);
1608 }
1609
1610 static int recurse_dir(struct thread_data *td, const char *dirname)
1611 {
1612         struct dirent *dir;
1613         int ret = 0;
1614         DIR *D;
1615
1616         D = opendir(dirname);
1617         if (!D) {
1618                 char buf[FIO_VERROR_SIZE];
1619
1620                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1621                 td_verror(td, errno, buf);
1622                 return 1;
1623         }
1624
1625         while ((dir = readdir(D)) != NULL) {
1626                 char full_path[PATH_MAX];
1627                 struct stat sb;
1628
1629                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1630                         continue;
1631
1632                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1633
1634                 if (lstat(full_path, &sb) == -1) {
1635                         if (errno != ENOENT) {
1636                                 td_verror(td, errno, "stat");
1637                                 ret = 1;
1638                                 break;
1639                         }
1640                 }
1641
1642                 if (S_ISREG(sb.st_mode)) {
1643                         add_file(td, full_path, 0, 1);
1644                         continue;
1645                 }
1646                 if (!S_ISDIR(sb.st_mode))
1647                         continue;
1648
1649                 ret = recurse_dir(td, full_path);
1650                 if (ret)
1651                         break;
1652         }
1653
1654         closedir(D);
1655         return ret;
1656 }
1657
1658 int add_dir_files(struct thread_data *td, const char *path)
1659 {
1660         int ret = recurse_dir(td, path);
1661
1662         if (!ret)
1663                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1664
1665         return ret;
1666 }
1667
1668 void dup_files(struct thread_data *td, struct thread_data *org)
1669 {
1670         struct fio_file *f;
1671         unsigned int i;
1672
1673         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1674
1675         if (!org->files)
1676                 return;
1677
1678         td->files = malloc(org->files_index * sizeof(f));
1679
1680         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1681                 td->file_locks = malloc(org->files_index);
1682
1683         for_each_file(org, f, i) {
1684                 struct fio_file *__f;
1685
1686                 __f = alloc_new_file(td);
1687
1688                 if (f->file_name) {
1689                         __f->file_name = smalloc_strdup(f->file_name);
1690                         if (!__f->file_name)
1691                                 assert(0);
1692
1693                         __f->filetype = f->filetype;
1694                 }
1695
1696                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1697                         __f->lock = f->lock;
1698                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1699                         __f->rwlock = f->rwlock;
1700
1701                 td->files[i] = __f;
1702         }
1703 }
1704
1705 /*
1706  * Returns the index that matches the filename, or -1 if not there
1707  */
1708 int get_fileno(struct thread_data *td, const char *fname)
1709 {
1710         struct fio_file *f;
1711         unsigned int i;
1712
1713         for_each_file(td, f, i)
1714                 if (!strcmp(f->file_name, fname))
1715                         return i;
1716
1717         return -1;
1718 }
1719
1720 /*
1721  * For log usage, where we add/open/close files automatically
1722  */
1723 void free_release_files(struct thread_data *td)
1724 {
1725         close_files(td);
1726         td->o.nr_files = 0;
1727         td->o.open_files = 0;
1728         td->files_index = 0;
1729         td->nr_normal_files = 0;
1730 }
1731
1732 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1733 {
1734         int i;
1735
1736         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1737                 f->last_pos[i] = f->file_offset;
1738                 f->last_start[i] = -1ULL;
1739         }
1740
1741         if (fio_file_axmap(f))
1742                 axmap_reset(f->io_axmap);
1743         else if (fio_file_lfsr(f))
1744                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1745 }
1746
1747 bool fio_files_done(struct thread_data *td)
1748 {
1749         struct fio_file *f;
1750         unsigned int i;
1751
1752         for_each_file(td, f, i)
1753                 if (!fio_file_done(f))
1754                         return false;
1755
1756         return true;
1757 }
1758
1759 /* free memory used in initialization phase only */
1760 void filesetup_mem_free(void)
1761 {
1762         free_already_allocated();
1763 }