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         left = f->real_file_size;
 163         bs = td->o.max_bs[DDIR_WRITE];
 164         if (bs > left)
 165                 bs = left;
 166
 167         b = malloc(bs);
 168         if (!b) {
 169                 td_verror(td, errno, "malloc");
 170                 goto err;
 171         }
 172
 173         while (left && !td->terminate) {
 174                 if (bs > left)
 175                         bs = left;
 176
 177                 fill_io_buffer(td, b, bs, bs);
 178
 179                 r = write(f->fd, b, bs);
 180
 181                 if (r > 0) {
 182                         left -= r;
 183                         continue;
 184                 } else {
 185                         if (r < 0) {
 186                                 int __e = errno;
 187
 188                                 if (__e == ENOSPC) {
 189                                         if (td->o.fill_device)
 190                                                 break;
 191                                         log_info("fio: ENOSPC on laying out "
 192                                                  "file, stopping\n");
 193                                         break;
 194                                 }
 195                                 td_verror(td, errno, "write");
 196                         } else
 197                                 td_verror(td, EIO, "write");
 198
 199                         break;
 200                 }
 201         }
 202
 203         if (td->terminate) {
 204                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 205                 td_io_unlink_file(td, f);
 206         } else if (td->o.create_fsync) {
 207                 if (fsync(f->fd) < 0) {
 208                         td_verror(td, errno, "fsync");
 209                         goto err;
 210                 }
 211         }
 212         if (td->o.fill_device && !td_write(td)) {
 213                 fio_file_clear_size_known(f);
 214                 if (td_io_get_file_size(td, f))
 215                         goto err;
 216                 if (f->io_size > f->real_file_size)
 217                         f->io_size = f->real_file_size;
 218         }
 219
 220         free(b);
 221 done:
 222         return 0;
 223 err:
 224         close(f->fd);
 225         f->fd = -1;
 226         if (b)
 227                 free(b);
 228         return 1;
 229 }
 230
 231 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 232 {
 233         int ret = 0, r, did_open = 0, old_runstate;
 234         unsigned long long left;
 235         unsigned int bs;
 236         char *b;
 237
 238         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 239             td_ioengine_flagged(td, FIO_NOIO))
 240                 return 0;
 241
 242         if (f->filetype == FIO_TYPE_CHAR)
 243                 return 0;
 244
 245         if (!fio_file_open(f)) {
 246                 if (td->io_ops->open_file(td, f)) {
 247                         log_err("fio: cannot pre-read, failed to open file\n");
 248                         return 1;
 249                 }
 250                 did_open = 1;
 251         }
 252
 253         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 254
 255         left = f->io_size;
 256         bs = td->o.max_bs[DDIR_READ];
 257         if (bs > left)
 258                 bs = left;
 259
 260         b = malloc(bs);
 261         if (!b) {
 262                 td_verror(td, errno, "malloc");
 263                 ret = 1;
 264                 goto error;
 265         }
 266         memset(b, 0, bs);
 267
 268         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 269                 td_verror(td, errno, "lseek");
 270                 log_err("fio: failed to lseek pre-read file\n");
 271                 ret = 1;
 272                 goto error;
 273         }
 274
 275         while (left && !td->terminate) {
 276                 if (bs > left)
 277                         bs = left;
 278
 279                 r = read(f->fd, b, bs);
 280
 281                 if (r == (int) bs) {
 282                         left -= bs;
 283                         continue;
 284                 } else {
 285                         td_verror(td, EIO, "pre_read");
 286                         break;
 287                 }
 288         }
 289
 290 error:
 291         td_restore_runstate(td, old_runstate);
 292
 293         if (did_open)
 294                 td->io_ops->close_file(td, f);
 295
 296         free(b);
 297         return ret;
 298 }
 299
 300 unsigned long long get_rand_file_size(struct thread_data *td)
 301 {
 302         unsigned long long ret, sized;
 303         uint64_t frand_max;
 304         unsigned long r;
 305
 306         frand_max = rand_max(&td->file_size_state);
 307         r = __rand(&td->file_size_state);
 308         sized = td->o.file_size_high - td->o.file_size_low;
 309         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 310         ret += td->o.file_size_low;
 311         ret -= (ret % td->o.rw_min_bs);
 312         return ret;
 313 }
 314
 315 static int file_size(struct thread_data *td, struct fio_file *f)
 316 {
 317         struct stat st;
 318
 319         if (stat(f->file_name, &st) == -1) {
 320                 td_verror(td, errno, "fstat");
 321                 return 1;
 322         }
 323
 324         f->real_file_size = st.st_size;
 325         return 0;
 326 }
 327
 328 static int bdev_size(struct thread_data *td, struct fio_file *f)
 329 {
 330         unsigned long long bytes = 0;
 331         int r;
 332
 333         if (td->io_ops->open_file(td, f)) {
 334                 log_err("fio: failed opening blockdev %s for size check\n",
 335                         f->file_name);
 336                 return 1;
 337         }
 338
 339         r = blockdev_size(f, &bytes);
 340         if (r) {
 341                 td_verror(td, r, "blockdev_size");
 342                 goto err;
 343         }
 344
 345         if (!bytes) {
 346                 log_err("%s: zero sized block device?\n", f->file_name);
 347                 goto err;
 348         }
 349
 350         f->real_file_size = bytes;
 351         td->io_ops->close_file(td, f);
 352         return 0;
 353 err:
 354         td->io_ops->close_file(td, f);
 355         return 1;
 356 }
 357
 358 static int char_size(struct thread_data *td, struct fio_file *f)
 359 {
 360 #ifdef FIO_HAVE_CHARDEV_SIZE
 361         unsigned long long bytes = 0;
 362         int r;
 363
 364         if (td->io_ops->open_file(td, f)) {
 365                 log_err("fio: failed opening chardev %s for size check\n",
 366                         f->file_name);
 367                 return 1;
 368         }
 369
 370         r = chardev_size(f, &bytes);
 371         if (r) {
 372                 td_verror(td, r, "chardev_size");
 373                 goto err;
 374         }
 375
 376         if (!bytes) {
 377                 log_err("%s: zero sized char device?\n", f->file_name);
 378                 goto err;
 379         }
 380
 381         f->real_file_size = bytes;
 382         td->io_ops->close_file(td, f);
 383         return 0;
 384 err:
 385         td->io_ops->close_file(td, f);
 386         return 1;
 387 #else
 388         f->real_file_size = -1ULL;
 389         return 0;
 390 #endif
 391 }
 392
 393 static int get_file_size(struct thread_data *td, struct fio_file *f)
 394 {
 395         int ret = 0;
 396
 397         if (fio_file_size_known(f))
 398                 return 0;
 399
 400         if (f->filetype == FIO_TYPE_FILE)
 401                 ret = file_size(td, f);
 402         else if (f->filetype == FIO_TYPE_BLOCK)
 403                 ret = bdev_size(td, f);
 404         else if (f->filetype == FIO_TYPE_CHAR)
 405                 ret = char_size(td, f);
 406         else
 407                 f->real_file_size = -1ULL;
 408
 409         /*
 410          * Leave ->real_file_size with 0 since it could be expectation
 411          * of initial setup for regular files.
 412          */
 413         if (ret)
 414                 return ret;
 415
 416         /*
 417          * If ->real_file_size is -1, a conditional for the message
 418          * "offset extends end" is always true, but it makes no sense,
 419          * so just return the same value here.
 420          */
 421         if (f->real_file_size == -1ULL) {
 422                 log_info("%s: failed to get file size of %s\n", td->o.name,
 423                                         f->file_name);
 424                 return 1;
 425         }
 426
 427         if (td->o.start_offset && f->file_offset == 0)
 428                 dprint(FD_FILE, "offset of file %s not initialized yet\n",
 429                                         f->file_name);
 430         /*
 431          * ->file_offset normally hasn't been initialized yet, so this
 432          * is basically always false.
 433          */
 434         if (f->file_offset > f->real_file_size) {
 435                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 436                                         (unsigned long long) f->file_offset,
 437                                         (unsigned long long) f->real_file_size);
 438                 return 1;
 439         }
 440
 441         fio_file_set_size_known(f);
 442         return 0;
 443 }
 444
 445 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 446                                    unsigned long long off,
 447                                    unsigned long long len)
 448 {
 449         int errval = 0, ret = 0;
 450
 451 #ifdef CONFIG_ESX
 452         return 0;
 453 #endif
 454
 455         if (len == -1ULL)
 456                 len = f->io_size;
 457         if (off == -1ULL)
 458                 off = f->file_offset;
 459
 460         if (len == -1ULL || off == -1ULL)
 461                 return 0;
 462
 463         if (td->io_ops->invalidate) {
 464                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 465                         f->file_name);
 466                 ret = td->io_ops->invalidate(td, f);
 467                 if (ret < 0)
 468                         errval = -ret;
 469         } else if (f->filetype == FIO_TYPE_FILE) {
 470                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 471                         f->file_name, off, len);
 472                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 473                 if (ret)
 474                         errval = ret;
 475         } else if (f->filetype == FIO_TYPE_BLOCK) {
 476                 int retry_count = 0;
 477
 478                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 479                 ret = blockdev_invalidate_cache(f);
 480                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 481                         /*
 482                          * Linux multipath devices reject ioctl while
 483                          * the maps are being updated. That window can
 484                          * last tens of milliseconds; we'll try up to
 485                          * a quarter of a second.
 486                          */
 487                         usleep(10000);
 488                         ret = blockdev_invalidate_cache(f);
 489                 }
 490                 if (ret < 0 && errno == EACCES && geteuid()) {
 491                         if (!root_warn) {
 492                                 log_err("fio: only root may flush block "
 493                                         "devices. Cache flush bypassed!\n");
 494                                 root_warn = 1;
 495                         }
 496                         ret = 0;
 497                 }
 498                 if (ret < 0)
 499                         errval = errno;
 500                 else if (ret) /* probably not supported */
 501                         errval = ret;
 502         } else if (f->filetype == FIO_TYPE_CHAR ||
 503                    f->filetype == FIO_TYPE_PIPE) {
 504                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 505                 ret = 0;
 506         }
 507
 508         /*
 509          * Cache flushing isn't a fatal condition, and we know it will
 510          * happen on some platforms where we don't have the proper
 511          * function to flush eg block device caches. So just warn and
 512          * continue on our way.
 513          */
 514         if (errval)
 515                 log_info("fio: cache invalidation of %s failed: %s\n",
 516                          f->file_name, strerror(errval));
 517
 518         return 0;
 519
 520 }
 521
 522 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 523 {
 524         if (!fio_file_open(f))
 525                 return 0;
 526
 527         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 528 }
 529
 530 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 531 {
 532         int ret = 0;
 533
 534         dprint(FD_FILE, "fd close %s\n", f->file_name);
 535
 536         remove_file_hash(f);
 537
 538         if (close(f->fd) < 0)
 539                 ret = errno;
 540
 541         f->fd = -1;
 542
 543         if (f->shadow_fd != -1) {
 544                 close(f->shadow_fd);
 545                 f->shadow_fd = -1;
 546         }
 547
 548         f->engine_pos = 0;
 549         return ret;
 550 }
 551
 552 int file_lookup_open(struct fio_file *f, int flags)
 553 {
 554         struct fio_file *__f;
 555         int from_hash;
 556
 557         __f = lookup_file_hash(f->file_name);
 558         if (__f) {
 559                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 560                 f->lock = __f->lock;
 561                 from_hash = 1;
 562         } else {
 563                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 564                 from_hash = 0;
 565         }
 566
 567 #ifdef WIN32
 568         flags |= _O_BINARY;
 569 #endif
 570
 571         f->fd = open(f->file_name, flags, 0600);
 572         return from_hash;
 573 }
 574
 575 static int file_close_shadow_fds(struct thread_data *td)
 576 {
 577         struct fio_file *f;
 578         int num_closed = 0;
 579         unsigned int i;
 580
 581         for_each_file(td, f, i) {
 582                 if (f->shadow_fd == -1)
 583                         continue;
 584
 585                 close(f->shadow_fd);
 586                 f->shadow_fd = -1;
 587                 num_closed++;
 588         }
 589
 590         return num_closed;
 591 }
 592
 593 int generic_open_file(struct thread_data *td, struct fio_file *f)
 594 {
 595         int is_std = 0;
 596         int flags = 0;
 597         int from_hash = 0;
 598
 599         dprint(FD_FILE, "fd open %s\n", f->file_name);
 600
 601         if (!strcmp(f->file_name, "-")) {
 602                 if (td_rw(td)) {
 603                         log_err("fio: can't read/write to stdin/out\n");
 604                         return 1;
 605                 }
 606                 is_std = 1;
 607
 608                 /*
 609                  * move output logging to stderr, if we are writing to stdout
 610                  */
 611                 if (td_write(td))
 612                         f_out = stderr;
 613         }
 614
 615         if (td_trim(td))
 616                 goto skip_flags;
 617         if (td->o.odirect)
 618                 flags |= OS_O_DIRECT;
 619         if (td->o.oatomic) {
 620                 if (!FIO_O_ATOMIC) {
 621                         td_verror(td, EINVAL, "OS does not support atomic IO");
 622                         return 1;
 623                 }
 624                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 625         }
 626         if (td->o.sync_io)
 627                 flags |= O_SYNC;
 628         if (td->o.create_on_open && td->o.allow_create)
 629                 flags |= O_CREAT;
 630 skip_flags:
 631         if (f->filetype != FIO_TYPE_FILE)
 632                 flags |= FIO_O_NOATIME;
 633
 634 open_again:
 635         if (td_write(td)) {
 636                 if (!read_only)
 637                         flags |= O_RDWR;
 638
 639                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 640                         flags |= O_CREAT;
 641
 642                 if (is_std)
 643                         f->fd = dup(STDOUT_FILENO);
 644                 else
 645                         from_hash = file_lookup_open(f, flags);
 646         } else if (td_read(td)) {
 647                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 648                         flags |= O_RDWR;
 649                 else
 650                         flags |= O_RDONLY;
 651
 652                 if (is_std)
 653                         f->fd = dup(STDIN_FILENO);
 654                 else
 655                         from_hash = file_lookup_open(f, flags);
 656         } else if (td_trim(td)) {
 657                 assert(!td_rw(td)); /* should have matched above */
 658                 flags |= O_RDWR;
 659                 from_hash = file_lookup_open(f, flags);
 660         }
 661
 662         if (f->fd == -1) {
 663                 char buf[FIO_VERROR_SIZE];
 664                 int __e = errno;
 665
 666                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 667                         flags &= ~FIO_O_NOATIME;
 668                         goto open_again;
 669                 }
 670                 if (__e == EMFILE && file_close_shadow_fds(td))
 671                         goto open_again;
 672
 673                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 674
 675                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 676                         log_err("fio: looks like your file system does not " \
 677                                 "support direct=1/buffered=0\n");
 678                 }
 679
 680                 td_verror(td, __e, buf);
 681                 return 1;
 682         }
 683
 684         if (!from_hash && f->fd != -1) {
 685                 if (add_file_hash(f)) {
 686                         int fio_unused ret;
 687
 688                         /*
 689                          * Stash away descriptor for later close. This is to
 690                          * work-around a "feature" on Linux, where a close of
 691                          * an fd that has been opened for write will trigger
 692                          * udev to call blkid to check partitions, fs id, etc.
 693                          * That pollutes the device cache, which can slow down
 694                          * unbuffered accesses.
 695                          */
 696                         if (f->shadow_fd == -1)
 697                                 f->shadow_fd = f->fd;
 698                         else {
 699                                 /*
 700                                  * OK to ignore, we haven't done anything
 701                                  * with it
 702                                  */
 703                                 ret = generic_close_file(td, f);
 704                         }
 705                         goto open_again;
 706                 }
 707         }
 708
 709         return 0;
 710 }
 711
 712 /*
 713  * This function i.e. get_file_size() is the default .get_file_size
 714  * implementation of majority of I/O engines.
 715  */
 716 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 717 {
 718         return get_file_size(td, f);
 719 }
 720
 721 /*
 722  * open/close all files, so that ->real_file_size gets set
 723  */
 724 static int get_file_sizes(struct thread_data *td)
 725 {
 726         struct fio_file *f;
 727         unsigned int i;
 728         int err = 0;
 729
 730         for_each_file(td, f, i) {
 731                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 732                                                                 f->file_name);
 733
 734                 if (td_io_get_file_size(td, f)) {
 735                         if (td->error != ENOENT) {
 736                                 log_err("%s\n", td->verror);
 737                                 err = 1;
 738                                 break;
 739                         }
 740                         clear_error(td);
 741                 }
 742
 743                 /*
 744                  * There are corner cases where we end up with -1 for
 745                  * ->real_file_size due to unsupported file type, etc.
 746                  * We then just set to size option value divided by number
 747                  * of files, similar to the way file ->io_size is set.
 748                  * stat(2) failure doesn't set ->real_file_size to -1.
 749                  */
 750                 if (f->real_file_size == -1ULL && td->o.size)
 751                         f->real_file_size = td->o.size / td->o.nr_files;
 752         }
 753
 754         return err;
 755 }
 756
 757 struct fio_mount {
 758         struct flist_head list;
 759         const char *base;
 760         char __base[256];
 761         unsigned int key;
 762 };
 763
 764 /*
 765  * Get free number of bytes for each file on each unique mount.
 766  */
 767 static unsigned long long get_fs_free_counts(struct thread_data *td)
 768 {
 769         struct flist_head *n, *tmp;
 770         unsigned long long ret = 0;
 771         struct fio_mount *fm;
 772         FLIST_HEAD(list);
 773         struct fio_file *f;
 774         unsigned int i;
 775
 776         for_each_file(td, f, i) {
 777                 struct stat sb;
 778                 char buf[256];
 779
 780                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 781                         if (f->real_file_size != -1ULL)
 782                                 ret += f->real_file_size;
 783                         continue;
 784                 } else if (f->filetype != FIO_TYPE_FILE)
 785                         continue;
 786
 787                 buf[255] = '\0';
 788                 strncpy(buf, f->file_name, 255);
 789
 790                 if (stat(buf, &sb) < 0) {
 791                         if (errno != ENOENT)
 792                                 break;
 793                         strcpy(buf, ".");
 794                         if (stat(buf, &sb) < 0)
 795                                 break;
 796                 }
 797
 798                 fm = NULL;
 799                 flist_for_each(n, &list) {
 800                         fm = flist_entry(n, struct fio_mount, list);
 801                         if (fm->key == sb.st_dev)
 802                                 break;
 803
 804                         fm = NULL;
 805                 }
 806
 807                 if (fm)
 808                         continue;
 809
 810                 fm = calloc(1, sizeof(*fm));
 811                 strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
 812                 fm->base = basename(fm->__base);
 813                 fm->key = sb.st_dev;
 814                 flist_add(&fm->list, &list);
 815         }
 816
 817         flist_for_each_safe(n, tmp, &list) {
 818                 unsigned long long sz;
 819
 820                 fm = flist_entry(n, struct fio_mount, list);
 821                 flist_del(&fm->list);
 822
 823                 sz = get_fs_free_size(fm->base);
 824                 if (sz && sz != -1ULL)
 825                         ret += sz;
 826
 827                 free(fm);
 828         }
 829
 830         return ret;
 831 }
 832
 833 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 834 {
 835         struct thread_options *o = &td->o;
 836         unsigned long long align_bs;  /* align the offset to this block size */
 837         unsigned long long offset;  /* align_bs-aligned offset */
 838
 839         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 840                 return f->real_file_size;
 841
 842         if (o->start_offset_percent > 0) {
 843
 844                 /* if blockalign is provided, find the min across read, write, and trim */
 845                 if (fio_option_is_set(o, ba)) {
 846                         align_bs = (unsigned long long) min(o->ba[DDIR_READ], o->ba[DDIR_WRITE]);
 847                         align_bs = min((unsigned long long) o->ba[DDIR_TRIM], align_bs);
 848                 } else {  /* else take the minimum block size */
 849                         align_bs = td_min_bs(td);
 850                 }
 851
 852                 /* calculate the raw offset */
 853                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 854                         (td->subjob_number * o->offset_increment);
 855
 856                 /* block align the offset at the next available boundary at
 857                    ceiling(offset / align_bs) * align_bs */
 858                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 859
 860         } else {  /* start_offset_percent not set */
 861                 offset = o->start_offset + o->start_offset +
 862                         td->subjob_number * o->offset_increment;
 863         }
 864         return offset;
 865 }
 866
 867 /*
 868  * Open the files and setup files sizes, creating files if necessary.
 869  */
 870 int setup_files(struct thread_data *td)
 871 {
 872         unsigned long long total_size, extend_size;
 873         struct thread_options *o = &td->o;
 874         struct fio_file *f;
 875         unsigned int i, nr_fs_extra = 0;
 876         int err = 0, need_extend;
 877         int old_state;
 878         const unsigned int bs = td_min_bs(td);
 879         uint64_t fs = 0;
 880
 881         dprint(FD_FILE, "setup files\n");
 882
 883         old_state = td_bump_runstate(td, TD_SETTING_UP);
 884
 885         if (o->read_iolog_file)
 886                 goto done;
 887
 888         /*
 889          * Find out physical size of files or devices for this thread,
 890          * before we determine I/O size and range of our targets.
 891          * If ioengine defines a setup() method, it's responsible for
 892          * opening the files and setting f->real_file_size to indicate
 893          * the valid range for that file.
 894          */
 895         if (td->io_ops->setup)
 896                 err = td->io_ops->setup(td);
 897         else
 898                 err = get_file_sizes(td);
 899
 900         if (err)
 901                 goto err_out;
 902
 903         /*
 904          * check sizes. if the files/devices do not exist and the size
 905          * isn't passed to fio, abort.
 906          */
 907         total_size = 0;
 908         for_each_file(td, f, i) {
 909                 f->fileno = i;
 910                 if (f->real_file_size == -1ULL)
 911                         total_size = -1ULL;
 912                 else
 913                         total_size += f->real_file_size;
 914         }
 915
 916         if (o->fill_device)
 917                 td->fill_device_size = get_fs_free_counts(td);
 918
 919         /*
 920          * device/file sizes are zero and no size given, punt
 921          */
 922         if ((!total_size || total_size == -1ULL) && !o->size &&
 923             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
 924             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 925                 log_err("%s: you need to specify size=\n", o->name);
 926                 td_verror(td, EINVAL, "total_file_size");
 927                 goto err_out;
 928         }
 929
 930         /*
 931          * Calculate per-file size and potential extra size for the
 932          * first files, if needed (i.e. if we don't have a fixed size).
 933          */
 934         if (!o->file_size_low && o->nr_files) {
 935                 uint64_t all_fs;
 936
 937                 fs = o->size / o->nr_files;
 938                 all_fs = fs * o->nr_files;
 939
 940                 if (all_fs < o->size)
 941                         nr_fs_extra = (o->size - all_fs) / bs;
 942         }
 943
 944         /*
 945          * now file sizes are known, so we can set ->io_size. if size= is
 946          * not given, ->io_size is just equal to ->real_file_size. if size
 947          * is given, ->io_size is size / nr_files.
 948          */
 949         extend_size = total_size = 0;
 950         need_extend = 0;
 951         for_each_file(td, f, i) {
 952                 f->file_offset = get_start_offset(td, f);
 953
 954                 /*
 955                  * Update ->io_size depending on options specified.
 956                  * ->file_size_low being 0 means filesize option isn't set.
 957                  * Non zero ->file_size_low equals ->file_size_high means
 958                  * filesize option is set in a fixed size format.
 959                  * Non zero ->file_size_low not equals ->file_size_high means
 960                  * filesize option is set in a range format.
 961                  */
 962                 if (!o->file_size_low) {
 963                         /*
 964                          * no file size or range given, file size is equal to
 965                          * total size divided by number of files. If the size
 966                          * doesn't divide nicely with the min blocksize,
 967                          * make the first files bigger.
 968                          */
 969                         f->io_size = fs;
 970                         if (nr_fs_extra) {
 971                                 nr_fs_extra--;
 972                                 f->io_size += bs;
 973                         }
 974
 975                         /*
 976                          * We normally don't come here for regular files, but
 977                          * if the result is 0 for a regular file, set it to the
 978                          * real file size. This could be size of the existing
 979                          * one if it already exists, but otherwise will be set
 980                          * to 0. A new file won't be created because
 981                          * ->io_size + ->file_offset equals ->real_file_size.
 982                          */
 983                         if (!f->io_size) {
 984                                 if (f->file_offset > f->real_file_size)
 985                                         goto err_offset;
 986                                 f->io_size = f->real_file_size - f->file_offset;
 987                                 if (!f->io_size)
 988                                         log_info("fio: file %s may be ignored\n",
 989                                                 f->file_name);
 990                         }
 991                 } else if (f->real_file_size < o->file_size_low ||
 992                            f->real_file_size > o->file_size_high) {
 993                         if (f->file_offset > o->file_size_low)
 994                                 goto err_offset;
 995                         /*
 996                          * file size given. if it's fixed, use that. if it's a
 997                          * range, generate a random size in-between.
 998                          */
 999                         if (o->file_size_low == o->file_size_high)
1000                                 f->io_size = o->file_size_low - f->file_offset;
1001                         else {
1002                                 f->io_size = get_rand_file_size(td)
1003                                                 - f->file_offset;
1004                         }
1005                 } else
1006                         f->io_size = f->real_file_size - f->file_offset;
1007
1008                 if (f->io_size == -1ULL)
1009                         total_size = -1ULL;
1010                 else {
1011                         if (o->size_percent) {
1012                                 uint64_t file_size;
1013
1014                                 file_size = f->io_size + f->file_offset;
1015                                 f->io_size = (file_size *
1016                                               o->size_percent) / 100;
1017                                 if (f->io_size > (file_size - f->file_offset))
1018                                         f->io_size = file_size - f->file_offset;
1019
1020                                 f->io_size -= (f->io_size % td_min_bs(td));
1021                         }
1022                         total_size += f->io_size;
1023                 }
1024
1025                 if (f->filetype == FIO_TYPE_FILE &&
1026                     (f->io_size + f->file_offset) > f->real_file_size &&
1027                     !td_ioengine_flagged(td, FIO_DISKLESSIO)) {
1028                         if (!o->create_on_open) {
1029                                 need_extend++;
1030                                 extend_size += (f->io_size + f->file_offset);
1031                                 fio_file_set_extend(f);
1032                         } else
1033                                 f->real_file_size = f->io_size + f->file_offset;
1034                 }
1035         }
1036
1037         if (td->o.block_error_hist) {
1038                 int len;
1039
1040                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1041                 f = td->files[0];
1042                 len = f->io_size / td->o.bs[DDIR_TRIM];
1043                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1044                         log_err("fio: cannot calculate block histogram with "
1045                                 "%d trim blocks, maximum %d\n",
1046                                 len, MAX_NR_BLOCK_INFOS);
1047                         td_verror(td, EINVAL, "block_error_hist");
1048                         goto err_out;
1049                 }
1050
1051                 td->ts.nr_block_infos = len;
1052                 for (i = 0; i < len; i++)
1053                         td->ts.block_infos[i] =
1054                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1055         } else
1056                 td->ts.nr_block_infos = 0;
1057
1058         if (!o->size || (total_size && o->size > total_size))
1059                 o->size = total_size;
1060
1061         if (o->size < td_min_bs(td)) {
1062                 log_err("fio: blocksize too large for data set\n");
1063                 goto err_out;
1064         }
1065
1066         /*
1067          * See if we need to extend some files, typically needed when our
1068          * target regular files don't exist yet, but our jobs require them
1069          * initially due to read I/Os.
1070          */
1071         if (need_extend) {
1072                 temp_stall_ts = 1;
1073                 if (output_format & FIO_OUTPUT_NORMAL) {
1074                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1075                                  o->name,
1076                                  need_extend > 1 ? "s" : "",
1077                                  need_extend,
1078                                  need_extend > 1 ? "s" : "",
1079                                  need_extend > 1 ? "total " : "",
1080                                  extend_size >> 20);
1081                 }
1082
1083                 for_each_file(td, f, i) {
1084                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1085
1086                         if (!fio_file_extend(f))
1087                                 continue;
1088
1089                         assert(f->filetype == FIO_TYPE_FILE);
1090                         fio_file_clear_extend(f);
1091                         if (!o->fill_device) {
1092                                 old_len = f->real_file_size;
1093                                 extend_len = f->io_size + f->file_offset -
1094                                                 old_len;
1095                         }
1096                         f->real_file_size = (f->io_size + f->file_offset);
1097                         err = extend_file(td, f);
1098                         if (err)
1099                                 break;
1100
1101                         err = __file_invalidate_cache(td, f, old_len,
1102                                                                 extend_len);
1103
1104                         /*
1105                          * Shut up static checker
1106                          */
1107                         if (f->fd != -1)
1108                                 close(f->fd);
1109
1110                         f->fd = -1;
1111                         if (err)
1112                                 break;
1113                 }
1114                 temp_stall_ts = 0;
1115         }
1116
1117         if (err)
1118                 goto err_out;
1119
1120         if (!o->zone_size)
1121                 o->zone_size = o->size;
1122
1123         /*
1124          * iolog already set the total io size, if we read back
1125          * stored entries.
1126          */
1127         if (!o->read_iolog_file) {
1128                 if (o->io_size)
1129                         td->total_io_size = o->io_size * o->loops;
1130                 else
1131                         td->total_io_size = o->size * o->loops;
1132         }
1133
1134 done:
1135         if (o->create_only)
1136                 td->done = 1;
1137
1138         td_restore_runstate(td, old_state);
1139         return 0;
1140 err_offset:
1141         log_err("%s: you need to specify valid offset=\n", o->name);
1142 err_out:
1143         td_restore_runstate(td, old_state);
1144         return 1;
1145 }
1146
1147 int pre_read_files(struct thread_data *td)
1148 {
1149         struct fio_file *f;
1150         unsigned int i;
1151
1152         dprint(FD_FILE, "pre_read files\n");
1153
1154         for_each_file(td, f, i) {
1155                 if (pre_read_file(td, f))
1156                         return -1;
1157         }
1158
1159         return 0;
1160 }
1161
1162 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1163 {
1164         unsigned int range_size, seed;
1165         unsigned long nranges;
1166         uint64_t fsize;
1167
1168         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1169         fsize = min(f->real_file_size, f->io_size);
1170
1171         nranges = (fsize + range_size - 1) / range_size;
1172
1173         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1174         if (!td->o.rand_repeatable)
1175                 seed = td->rand_seeds[4];
1176
1177         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1178                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
1179         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1180                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
1181         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1182                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, seed);
1183
1184         return 1;
1185 }
1186
1187 static int init_rand_distribution(struct thread_data *td)
1188 {
1189         struct fio_file *f;
1190         unsigned int i;
1191         int state;
1192
1193         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
1194                 return 0;
1195
1196         state = td_bump_runstate(td, TD_SETTING_UP);
1197
1198         for_each_file(td, f, i)
1199                 __init_rand_distribution(td, f);
1200
1201         td_restore_runstate(td, state);
1202
1203         return 1;
1204 }
1205
1206 /*
1207  * Check if the number of blocks exceeds the randomness capability of
1208  * the selected generator. Tausworthe is 32-bit, the others are fullly
1209  * 64-bit capable.
1210  */
1211 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1212                                  uint64_t blocks)
1213 {
1214         if (blocks <= FRAND32_MAX)
1215                 return 0;
1216         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1217                 return 0;
1218
1219         /*
1220          * If the user hasn't specified a random generator, switch
1221          * to tausworthe64 with informational warning. If the user did
1222          * specify one, just warn.
1223          */
1224         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1225                         f->file_name);
1226
1227         if (!fio_option_is_set(&td->o, random_generator)) {
1228                 log_info("fio: Switching to tausworthe64. Use the "
1229                          "random_generator= option to get rid of this "
1230                          "warning.\n");
1231                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1232                 return 0;
1233         }
1234
1235         /*
1236          * Just make this information to avoid breaking scripts.
1237          */
1238         log_info("fio: Use the random_generator= option to switch to lfsr or "
1239                          "tausworthe64.\n");
1240         return 0;
1241 }
1242
1243 int init_random_map(struct thread_data *td)
1244 {
1245         unsigned long long blocks;
1246         struct fio_file *f;
1247         unsigned int i;
1248
1249         if (init_rand_distribution(td))
1250                 return 0;
1251         if (!td_random(td))
1252                 return 0;
1253
1254         for_each_file(td, f, i) {
1255                 uint64_t fsize = min(f->real_file_size, f->io_size);
1256
1257                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1258
1259                 if (check_rand_gen_limits(td, f, blocks))
1260                         return 1;
1261
1262                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1263                         unsigned long seed;
1264
1265                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1266
1267                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1268                                 fio_file_set_lfsr(f);
1269                                 continue;
1270                         }
1271                 } else if (!td->o.norandommap) {
1272                         f->io_axmap = axmap_new(blocks);
1273                         if (f->io_axmap) {
1274                                 fio_file_set_axmap(f);
1275                                 continue;
1276                         }
1277                 } else if (td->o.norandommap)
1278                         continue;
1279
1280                 if (!td->o.softrandommap) {
1281                         log_err("fio: failed allocating random map. If running"
1282                                 " a large number of jobs, try the 'norandommap'"
1283                                 " option or set 'softrandommap'. Or give"
1284                                 " a larger --alloc-size to fio.\n");
1285                         return 1;
1286                 }
1287
1288                 log_info("fio: file %s failed allocating random map. Running "
1289                          "job without.\n", f->file_name);
1290         }
1291
1292         return 0;
1293 }
1294
1295 void close_files(struct thread_data *td)
1296 {
1297         struct fio_file *f;
1298         unsigned int i;
1299
1300         for_each_file(td, f, i) {
1301                 if (fio_file_open(f))
1302                         td_io_close_file(td, f);
1303         }
1304 }
1305
1306 void close_and_free_files(struct thread_data *td)
1307 {
1308         struct fio_file *f;
1309         unsigned int i;
1310
1311         dprint(FD_FILE, "close files\n");
1312
1313         for_each_file(td, f, i) {
1314                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1315                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1316                         td_io_unlink_file(td, f);
1317                 }
1318
1319                 if (fio_file_open(f))
1320                         td_io_close_file(td, f);
1321
1322                 remove_file_hash(f);
1323
1324                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1325                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1326                         td_io_unlink_file(td, f);
1327                 }
1328
1329                 sfree(f->file_name);
1330                 f->file_name = NULL;
1331                 if (fio_file_axmap(f)) {
1332                         axmap_free(f->io_axmap);
1333                         f->io_axmap = NULL;
1334                 }
1335                 sfree(f);
1336         }
1337
1338         td->o.filename = NULL;
1339         free(td->files);
1340         free(td->file_locks);
1341         td->files_index = 0;
1342         td->files = NULL;
1343         td->file_locks = NULL;
1344         td->o.file_lock_mode = FILE_LOCK_NONE;
1345         td->o.nr_files = 0;
1346 }
1347
1348 static void get_file_type(struct fio_file *f)
1349 {
1350         struct stat sb;
1351
1352         if (!strcmp(f->file_name, "-"))
1353                 f->filetype = FIO_TYPE_PIPE;
1354         else
1355                 f->filetype = FIO_TYPE_FILE;
1356
1357 #ifdef WIN32
1358         /* \\.\ is the device namespace in Windows, where every file is
1359          * a block device */
1360         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1361                 f->filetype = FIO_TYPE_BLOCK;
1362 #endif
1363
1364         if (!stat(f->file_name, &sb)) {
1365                 if (S_ISBLK(sb.st_mode))
1366                         f->filetype = FIO_TYPE_BLOCK;
1367                 else if (S_ISCHR(sb.st_mode))
1368                         f->filetype = FIO_TYPE_CHAR;
1369                 else if (S_ISFIFO(sb.st_mode))
1370                         f->filetype = FIO_TYPE_PIPE;
1371         }
1372 }
1373
1374 static bool __is_already_allocated(const char *fname, bool set)
1375 {
1376         struct flist_head *entry;
1377         bool ret;
1378
1379         ret = file_bloom_exists(fname, set);
1380         if (!ret)
1381                 return ret;
1382
1383         flist_for_each(entry, &filename_list) {
1384                 struct file_name *fn;
1385
1386                 fn = flist_entry(entry, struct file_name, list);
1387
1388                 if (!strcmp(fn->filename, fname))
1389                         return true;
1390         }
1391
1392         return false;
1393 }
1394
1395 static bool is_already_allocated(const char *fname)
1396 {
1397         bool ret;
1398
1399         fio_file_hash_lock();
1400         ret = __is_already_allocated(fname, false);
1401         fio_file_hash_unlock();
1402
1403         return ret;
1404 }
1405
1406 static void set_already_allocated(const char *fname)
1407 {
1408         struct file_name *fn;
1409
1410         fn = malloc(sizeof(struct file_name));
1411         fn->filename = strdup(fname);
1412
1413         fio_file_hash_lock();
1414         if (!__is_already_allocated(fname, true)) {
1415                 flist_add_tail(&fn->list, &filename_list);
1416                 fn = NULL;
1417         }
1418         fio_file_hash_unlock();
1419
1420         if (fn) {
1421                 free(fn->filename);
1422                 free(fn);
1423         }
1424 }
1425
1426 static void free_already_allocated(void)
1427 {
1428         struct flist_head *entry, *tmp;
1429         struct file_name *fn;
1430
1431         if (flist_empty(&filename_list))
1432                 return;
1433
1434         fio_file_hash_lock();
1435         flist_for_each_safe(entry, tmp, &filename_list) {
1436                 fn = flist_entry(entry, struct file_name, list);
1437                 free(fn->filename);
1438                 flist_del(&fn->list);
1439                 free(fn);
1440         }
1441
1442         fio_file_hash_unlock();
1443 }
1444
1445 static struct fio_file *alloc_new_file(struct thread_data *td)
1446 {
1447         struct fio_file *f;
1448
1449         f = smalloc(sizeof(*f));
1450         if (!f) {
1451                 assert(0);
1452                 return NULL;
1453         }
1454
1455         f->fd = -1;
1456         f->shadow_fd = -1;
1457         fio_file_reset(td, f);
1458         return f;
1459 }
1460
1461 bool exists_and_not_regfile(const char *filename)
1462 {
1463         struct stat sb;
1464
1465         if (lstat(filename, &sb) == -1)
1466                 return false;
1467
1468 #ifndef WIN32 /* NOT Windows */
1469         if (S_ISREG(sb.st_mode))
1470                 return false;
1471 #else
1472         /* \\.\ is the device namespace in Windows, where every file
1473          * is a device node */
1474         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1475                 return false;
1476 #endif
1477
1478         return true;
1479 }
1480
1481 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1482 {
1483         int cur_files = td->files_index;
1484         char file_name[PATH_MAX];
1485         struct fio_file *f;
1486         int len = 0;
1487
1488         dprint(FD_FILE, "add file %s\n", fname);
1489
1490         if (td->o.directory)
1491                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1492                                         td->o.unique_filename);
1493
1494         sprintf(file_name + len, "%s", fname);
1495
1496         /* clean cloned siblings using existing files */
1497         if (numjob && is_already_allocated(file_name) &&
1498             !exists_and_not_regfile(fname))
1499                 return 0;
1500
1501         f = alloc_new_file(td);
1502
1503         if (td->files_size <= td->files_index) {
1504                 unsigned int new_size = td->o.nr_files + 1;
1505
1506                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1507
1508                 td->files = realloc(td->files, new_size * sizeof(f));
1509                 if (td->files == NULL) {
1510                         log_err("fio: realloc OOM\n");
1511                         assert(0);
1512                 }
1513                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1514                         td->file_locks = realloc(td->file_locks, new_size);
1515                         if (!td->file_locks) {
1516                                 log_err("fio: realloc OOM\n");
1517                                 assert(0);
1518                         }
1519                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1520                 }
1521                 td->files_size = new_size;
1522         }
1523         td->files[cur_files] = f;
1524         f->fileno = cur_files;
1525
1526         /*
1527          * init function, io engine may not be loaded yet
1528          */
1529         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1530                 f->real_file_size = -1ULL;
1531
1532         f->file_name = smalloc_strdup(file_name);
1533         if (!f->file_name)
1534                 assert(0);
1535
1536         get_file_type(f);
1537
1538         switch (td->o.file_lock_mode) {
1539         case FILE_LOCK_NONE:
1540                 break;
1541         case FILE_LOCK_READWRITE:
1542                 f->rwlock = fio_rwlock_init();
1543                 break;
1544         case FILE_LOCK_EXCLUSIVE:
1545                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1546                 break;
1547         default:
1548                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1549                 assert(0);
1550         }
1551
1552         td->files_index++;
1553         if (f->filetype == FIO_TYPE_FILE)
1554                 td->nr_normal_files++;
1555
1556         set_already_allocated(file_name);
1557
1558         if (inc)
1559                 td->o.nr_files++;
1560
1561         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1562                                                         cur_files);
1563
1564         return cur_files;
1565 }
1566
1567 int add_file_exclusive(struct thread_data *td, const char *fname)
1568 {
1569         struct fio_file *f;
1570         unsigned int i;
1571
1572         for_each_file(td, f, i) {
1573                 if (!strcmp(f->file_name, fname))
1574                         return i;
1575         }
1576
1577         return add_file(td, fname, 0, 1);
1578 }
1579
1580 void get_file(struct fio_file *f)
1581 {
1582         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1583         assert(fio_file_open(f));
1584         f->references++;
1585 }
1586
1587 int put_file(struct thread_data *td, struct fio_file *f)
1588 {
1589         int f_ret = 0, ret = 0;
1590
1591         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1592
1593         if (!fio_file_open(f)) {
1594                 assert(f->fd == -1);
1595                 return 0;
1596         }
1597
1598         assert(f->references);
1599         if (--f->references)
1600                 return 0;
1601
1602         if (should_fsync(td) && td->o.fsync_on_close) {
1603                 f_ret = fsync(f->fd);
1604                 if (f_ret < 0)
1605                         f_ret = errno;
1606         }
1607
1608         if (td->io_ops->close_file)
1609                 ret = td->io_ops->close_file(td, f);
1610
1611         if (!ret)
1612                 ret = f_ret;
1613
1614         td->nr_open_files--;
1615         fio_file_clear_open(f);
1616         assert(f->fd == -1);
1617         return ret;
1618 }
1619
1620 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1621 {
1622         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1623                 return;
1624
1625         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1626                 if (ddir == DDIR_READ)
1627                         fio_rwlock_read(f->rwlock);
1628                 else
1629                         fio_rwlock_write(f->rwlock);
1630         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1631                 fio_mutex_down(f->lock);
1632
1633         td->file_locks[f->fileno] = td->o.file_lock_mode;
1634 }
1635
1636 void unlock_file(struct thread_data *td, struct fio_file *f)
1637 {
1638         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1639                 return;
1640
1641         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1642                 fio_rwlock_unlock(f->rwlock);
1643         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1644                 fio_mutex_up(f->lock);
1645
1646         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1647 }
1648
1649 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1650 {
1651         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1652                 return;
1653         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1654                 unlock_file(td, f);
1655 }
1656
1657 static int recurse_dir(struct thread_data *td, const char *dirname)
1658 {
1659         struct dirent *dir;
1660         int ret = 0;
1661         DIR *D;
1662
1663         D = opendir(dirname);
1664         if (!D) {
1665                 char buf[FIO_VERROR_SIZE];
1666
1667                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1668                 td_verror(td, errno, buf);
1669                 return 1;
1670         }
1671
1672         while ((dir = readdir(D)) != NULL) {
1673                 char full_path[PATH_MAX];
1674                 struct stat sb;
1675
1676                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1677                         continue;
1678
1679                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1680
1681                 if (lstat(full_path, &sb) == -1) {
1682                         if (errno != ENOENT) {
1683                                 td_verror(td, errno, "stat");
1684                                 ret = 1;
1685                                 break;
1686                         }
1687                 }
1688
1689                 if (S_ISREG(sb.st_mode)) {
1690                         add_file(td, full_path, 0, 1);
1691                         continue;
1692                 }
1693                 if (!S_ISDIR(sb.st_mode))
1694                         continue;
1695
1696                 ret = recurse_dir(td, full_path);
1697                 if (ret)
1698                         break;
1699         }
1700
1701         closedir(D);
1702         return ret;
1703 }
1704
1705 int add_dir_files(struct thread_data *td, const char *path)
1706 {
1707         int ret = recurse_dir(td, path);
1708
1709         if (!ret)
1710                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1711
1712         return ret;
1713 }
1714
1715 void dup_files(struct thread_data *td, struct thread_data *org)
1716 {
1717         struct fio_file *f;
1718         unsigned int i;
1719
1720         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1721
1722         if (!org->files)
1723                 return;
1724
1725         td->files = malloc(org->files_index * sizeof(f));
1726
1727         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1728                 td->file_locks = malloc(org->files_index);
1729
1730         for_each_file(org, f, i) {
1731                 struct fio_file *__f;
1732
1733                 __f = alloc_new_file(td);
1734
1735                 if (f->file_name) {
1736                         __f->file_name = smalloc_strdup(f->file_name);
1737                         if (!__f->file_name)
1738                                 assert(0);
1739
1740                         __f->filetype = f->filetype;
1741                 }
1742
1743                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1744                         __f->lock = f->lock;
1745                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1746                         __f->rwlock = f->rwlock;
1747
1748                 td->files[i] = __f;
1749         }
1750 }
1751
1752 /*
1753  * Returns the index that matches the filename, or -1 if not there
1754  */
1755 int get_fileno(struct thread_data *td, const char *fname)
1756 {
1757         struct fio_file *f;
1758         unsigned int i;
1759
1760         for_each_file(td, f, i)
1761                 if (!strcmp(f->file_name, fname))
1762                         return i;
1763
1764         return -1;
1765 }
1766
1767 /*
1768  * For log usage, where we add/open/close files automatically
1769  */
1770 void free_release_files(struct thread_data *td)
1771 {
1772         close_files(td);
1773         td->o.nr_files = 0;
1774         td->o.open_files = 0;
1775         td->files_index = 0;
1776         td->nr_normal_files = 0;
1777 }
1778
1779 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1780 {
1781         int i;
1782
1783         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1784                 f->last_pos[i] = f->file_offset;
1785                 f->last_start[i] = -1ULL;
1786         }
1787
1788         if (fio_file_axmap(f))
1789                 axmap_reset(f->io_axmap);
1790         else if (fio_file_lfsr(f))
1791                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1792 }
1793
1794 bool fio_files_done(struct thread_data *td)
1795 {
1796         struct fio_file *f;
1797         unsigned int i;
1798
1799         for_each_file(td, f, i)
1800                 if (!fio_file_done(f))
1801                         return false;
1802
1803         return true;
1804 }
1805
1806 /* free memory used in initialization phase only */
1807 void filesetup_mem_free(void)
1808 {
1809         free_already_allocated();
1810 }