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
   9 #include "fio.h"
  10 #include "smalloc.h"
  11 #include "filehash.h"
  12 #include "options.h"
  13 #include "os/os.h"
  14 #include "hash.h"
  15 #include "lib/axmap.h"
  16 #include "rwlock.h"
  17 #include "zbd.h"
  18
  19 #ifdef CONFIG_LINUX_FALLOCATE
  20 #include <linux/falloc.h>
  21 #endif
  22
  23 static FLIST_HEAD(filename_list);
  24
  25 /*
  26  * List entry for filename_list
  27  */
  28 struct file_name {
  29         struct flist_head list;
  30         char *filename;
  31 };
  32
  33 static inline void clear_error(struct thread_data *td)
  34 {
  35         td->error = 0;
  36         td->verror[0] = '\0';
  37 }
  38
  39 static int native_fallocate(struct thread_data *td, struct fio_file *f)
  40 {
  41         bool success;
  42
  43         success = fio_fallocate(f, 0, f->real_file_size);
  44         dprint(FD_FILE, "native fallocate of file %s size %llu was "
  45                         "%ssuccessful\n", f->file_name,
  46                         (unsigned long long) f->real_file_size,
  47                         !success ? "un": "");
  48
  49         if (success)
  50                 return false;
  51
  52         if (errno == ENOSYS)
  53                 dprint(FD_FILE, "native fallocate is not implemented\n");
  54
  55         return true;
  56 }
  57
  58 static void fallocate_file(struct thread_data *td, struct fio_file *f)
  59 {
  60         if (td->o.fill_device)
  61                 return;
  62
  63         switch (td->o.fallocate_mode) {
  64         case FIO_FALLOCATE_NATIVE:
  65                 native_fallocate(td, f);
  66                 break;
  67         case FIO_FALLOCATE_NONE:
  68                 break;
  69 #ifdef CONFIG_POSIX_FALLOCATE
  70         case FIO_FALLOCATE_POSIX: {
  71                 int r;
  72
  73                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  74                                  f->file_name,
  75                                  (unsigned long long) f->real_file_size);
  76
  77                 r = posix_fallocate(f->fd, 0, f->real_file_size);
  78                 if (r > 0)
  79                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
  80                 break;
  81                 }
  82 #endif /* CONFIG_POSIX_FALLOCATE */
  83 #ifdef CONFIG_LINUX_FALLOCATE
  84         case FIO_FALLOCATE_KEEP_SIZE: {
  85                 int r;
  86
  87                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
  88                                 "file %s size %llu\n", f->file_name,
  89                                 (unsigned long long) f->real_file_size);
  90
  91                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
  92                 if (r != 0)
  93                         td_verror(td, errno, "fallocate");
  94
  95                 break;
  96                 }
  97 #endif /* CONFIG_LINUX_FALLOCATE */
  98         case FIO_FALLOCATE_TRUNCATE: {
  99                 int r;
 100
 101                 dprint(FD_FILE, "ftruncate file %s size %llu\n",
 102                                 f->file_name,
 103                                 (unsigned long long) f->real_file_size);
 104                 r = ftruncate(f->fd, f->real_file_size);
 105                 if (r != 0)
 106                         td_verror(td, errno, "ftruncate");
 107
 108                 break;
 109         }
 110         default:
 111                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
 112                 assert(0);
 113         }
 114 }
 115
 116 /*
 117  * Leaves f->fd open on success, caller must close
 118  */
 119 static int extend_file(struct thread_data *td, struct fio_file *f)
 120 {
 121         int new_layout = 0, unlink_file = 0, flags;
 122         unsigned long long left;
 123         unsigned long long bs;
 124         char *b = NULL;
 125
 126         if (read_only) {
 127                 log_err("fio: refusing extend of file due to read-only\n");
 128                 return 0;
 129         }
 130
 131         /*
 132          * check if we need to lay the file out complete again. fio
 133          * does that for operations involving reads, or for writes
 134          * where overwrite is set
 135          */
 136         if (td_read(td) ||
 137            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
 138             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
 139                 new_layout = 1;
 140         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
 141                 unlink_file = 1;
 142
 143         if (unlink_file || new_layout) {
 144                 int ret;
 145
 146                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
 147
 148                 ret = td_io_unlink_file(td, f);
 149                 if (ret != 0 && ret != ENOENT) {
 150                         td_verror(td, errno, "unlink");
 151                         return 1;
 152                 }
 153         }
 154
 155         flags = O_WRONLY;
 156         if (td->o.allow_create)
 157                 flags |= O_CREAT;
 158         if (new_layout)
 159                 flags |= O_TRUNC;
 160
 161 #ifdef WIN32
 162         flags |= _O_BINARY;
 163 #endif
 164
 165         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 166         f->fd = open(f->file_name, flags, 0644);
 167         if (f->fd < 0) {
 168                 int err = errno;
 169
 170                 if (err == ENOENT && !td->o.allow_create)
 171                         log_err("fio: file creation disallowed by "
 172                                         "allow_file_create=0\n");
 173                 else
 174                         td_verror(td, err, "open");
 175                 return 1;
 176         }
 177
 178         fallocate_file(td, f);
 179
 180         /*
 181          * If our jobs don't require regular files initially, we're done.
 182          */
 183         if (!new_layout)
 184                 goto done;
 185
 186         /*
 187          * The size will be -1ULL when fill_device is used, so don't truncate
 188          * or fallocate this file, just write it
 189          */
 190         if (!td->o.fill_device) {
 191                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 192                                         (unsigned long long) f->real_file_size);
 193                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 194                         if (errno != EFBIG) {
 195                                 td_verror(td, errno, "ftruncate");
 196                                 goto err;
 197                         }
 198                 }
 199         }
 200
 201         left = f->real_file_size;
 202         bs = td->o.max_bs[DDIR_WRITE];
 203         if (bs > left)
 204                 bs = left;
 205
 206         b = malloc(bs);
 207         if (!b) {
 208                 td_verror(td, errno, "malloc");
 209                 goto err;
 210         }
 211
 212         while (left && !td->terminate) {
 213                 ssize_t r;
 214
 215                 if (bs > left)
 216                         bs = left;
 217
 218                 fill_io_buffer(td, b, bs, bs);
 219
 220                 r = write(f->fd, b, bs);
 221
 222                 if (r > 0) {
 223                         left -= r;
 224                         continue;
 225                 } else {
 226                         if (r < 0) {
 227                                 int __e = errno;
 228
 229                                 if (__e == ENOSPC) {
 230                                         if (td->o.fill_device)
 231                                                 break;
 232                                         log_info("fio: ENOSPC on laying out "
 233                                                  "file, stopping\n");
 234                                 }
 235                                 td_verror(td, errno, "write");
 236                         } else
 237                                 td_verror(td, EIO, "write");
 238
 239                         goto err;
 240                 }
 241         }
 242
 243         if (td->terminate) {
 244                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 245                 td_io_unlink_file(td, f);
 246         } else if (td->o.create_fsync) {
 247                 if (fsync(f->fd) < 0) {
 248                         td_verror(td, errno, "fsync");
 249                         goto err;
 250                 }
 251         }
 252         if (td->o.fill_device && !td_write(td)) {
 253                 fio_file_clear_size_known(f);
 254                 if (td_io_get_file_size(td, f))
 255                         goto err;
 256                 if (f->io_size > f->real_file_size)
 257                         f->io_size = f->real_file_size;
 258         }
 259
 260         free(b);
 261 done:
 262         return 0;
 263 err:
 264         close(f->fd);
 265         f->fd = -1;
 266         if (b)
 267                 free(b);
 268         return 1;
 269 }
 270
 271 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 272 {
 273         int r, did_open = 0, old_runstate;
 274         unsigned long long left;
 275         unsigned long long bs;
 276         bool ret = true;
 277         char *b;
 278
 279         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 280             td_ioengine_flagged(td, FIO_NOIO))
 281                 return true;
 282
 283         if (f->filetype == FIO_TYPE_CHAR)
 284                 return true;
 285
 286         if (!fio_file_open(f)) {
 287                 if (td->io_ops->open_file(td, f)) {
 288                         log_err("fio: cannot pre-read, failed to open file\n");
 289                         return false;
 290                 }
 291                 did_open = 1;
 292         }
 293
 294         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 295
 296         left = f->io_size;
 297         bs = td->o.max_bs[DDIR_READ];
 298         if (bs > left)
 299                 bs = left;
 300
 301         b = malloc(bs);
 302         if (!b) {
 303                 td_verror(td, errno, "malloc");
 304                 ret = false;
 305                 goto error;
 306         }
 307         memset(b, 0, bs);
 308
 309         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 310                 td_verror(td, errno, "lseek");
 311                 log_err("fio: failed to lseek pre-read file\n");
 312                 ret = false;
 313                 goto error;
 314         }
 315
 316         while (left && !td->terminate) {
 317                 if (bs > left)
 318                         bs = left;
 319
 320                 r = read(f->fd, b, bs);
 321
 322                 if (r == (int) bs) {
 323                         left -= bs;
 324                         continue;
 325                 } else {
 326                         td_verror(td, EIO, "pre_read");
 327                         break;
 328                 }
 329         }
 330
 331 error:
 332         td_restore_runstate(td, old_runstate);
 333
 334         if (did_open)
 335                 td->io_ops->close_file(td, f);
 336
 337         free(b);
 338         return ret;
 339 }
 340
 341 unsigned long long get_rand_file_size(struct thread_data *td)
 342 {
 343         unsigned long long ret, sized;
 344         uint64_t frand_max;
 345         uint64_t r;
 346
 347         frand_max = rand_max(&td->file_size_state);
 348         r = __rand(&td->file_size_state);
 349         sized = td->o.file_size_high - td->o.file_size_low;
 350         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 351         ret += td->o.file_size_low;
 352         ret -= (ret % td->o.rw_min_bs);
 353         return ret;
 354 }
 355
 356 static int file_size(struct thread_data *td, struct fio_file *f)
 357 {
 358         struct stat st;
 359
 360         if (stat(f->file_name, &st) == -1) {
 361                 td_verror(td, errno, "fstat");
 362                 return 1;
 363         }
 364
 365         f->real_file_size = st.st_size;
 366         return 0;
 367 }
 368
 369 static int bdev_size(struct thread_data *td, struct fio_file *f)
 370 {
 371         unsigned long long bytes = 0;
 372         int r;
 373
 374         if (td->io_ops->open_file(td, f)) {
 375                 log_err("fio: failed opening blockdev %s for size check\n",
 376                         f->file_name);
 377                 return 1;
 378         }
 379
 380         r = blockdev_size(f, &bytes);
 381         if (r) {
 382                 td_verror(td, r, "blockdev_size");
 383                 goto err;
 384         }
 385
 386         if (!bytes) {
 387                 log_err("%s: zero sized block device?\n", f->file_name);
 388                 goto err;
 389         }
 390
 391         f->real_file_size = bytes;
 392         td->io_ops->close_file(td, f);
 393         return 0;
 394 err:
 395         td->io_ops->close_file(td, f);
 396         return 1;
 397 }
 398
 399 static int char_size(struct thread_data *td, struct fio_file *f)
 400 {
 401 #ifdef FIO_HAVE_CHARDEV_SIZE
 402         unsigned long long bytes = 0;
 403         int r;
 404
 405         if (td->io_ops->open_file(td, f)) {
 406                 log_err("fio: failed opening chardev %s for size check\n",
 407                         f->file_name);
 408                 return 1;
 409         }
 410
 411         r = chardev_size(f, &bytes);
 412         if (r) {
 413                 td_verror(td, r, "chardev_size");
 414                 goto err;
 415         }
 416
 417         if (!bytes) {
 418                 log_err("%s: zero sized char device?\n", f->file_name);
 419                 goto err;
 420         }
 421
 422         f->real_file_size = bytes;
 423         td->io_ops->close_file(td, f);
 424         return 0;
 425 err:
 426         td->io_ops->close_file(td, f);
 427         return 1;
 428 #else
 429         f->real_file_size = -1ULL;
 430         return 0;
 431 #endif
 432 }
 433
 434 static int get_file_size(struct thread_data *td, struct fio_file *f)
 435 {
 436         int ret = 0;
 437
 438         if (fio_file_size_known(f))
 439                 return 0;
 440
 441         if (f->filetype == FIO_TYPE_FILE)
 442                 ret = file_size(td, f);
 443         else if (f->filetype == FIO_TYPE_BLOCK)
 444                 ret = bdev_size(td, f);
 445         else if (f->filetype == FIO_TYPE_CHAR)
 446                 ret = char_size(td, f);
 447         else {
 448                 f->real_file_size = -1;
 449                 log_info("%s: failed to get file size of %s\n", td->o.name,
 450                                         f->file_name);
 451                 return 1; /* avoid offset extends end error message */
 452         }
 453
 454         /*
 455          * Leave ->real_file_size with 0 since it could be expectation
 456          * of initial setup for regular files.
 457          */
 458         if (ret)
 459                 return ret;
 460
 461         /*
 462          * ->file_offset normally hasn't been initialized yet, so this
 463          * is basically always false unless ->real_file_size is -1, but
 464          * if ->real_file_size is -1 this message doesn't make sense.
 465          * As a result, this message is basically useless.
 466          */
 467         if (f->file_offset > f->real_file_size) {
 468                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 469                                         (unsigned long long) f->file_offset,
 470                                         (unsigned long long) f->real_file_size);
 471                 return 1;
 472         }
 473
 474         fio_file_set_size_known(f);
 475         return 0;
 476 }
 477
 478 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 479                                    unsigned long long off,
 480                                    unsigned long long len)
 481 {
 482         int errval = 0, ret = 0;
 483
 484 #ifdef CONFIG_ESX
 485         return 0;
 486 #endif
 487
 488         if (len == -1ULL)
 489                 len = f->io_size;
 490         if (off == -1ULL)
 491                 off = f->file_offset;
 492
 493         if (len == -1ULL || off == -1ULL)
 494                 return 0;
 495
 496         if (td->io_ops->invalidate) {
 497                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 498                         f->file_name);
 499                 ret = td->io_ops->invalidate(td, f);
 500                 if (ret < 0)
 501                         errval = -ret;
 502         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 503                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
 504                        td->io_ops->name);
 505         } else if (f->filetype == FIO_TYPE_FILE) {
 506                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 507                         f->file_name, off, len);
 508                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 509                 if (ret)
 510                         errval = ret;
 511         } else if (f->filetype == FIO_TYPE_BLOCK) {
 512                 int retry_count = 0;
 513
 514                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 515                 ret = blockdev_invalidate_cache(f);
 516                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 517                         /*
 518                          * Linux multipath devices reject ioctl while
 519                          * the maps are being updated. That window can
 520                          * last tens of milliseconds; we'll try up to
 521                          * a quarter of a second.
 522                          */
 523                         usleep(10000);
 524                         ret = blockdev_invalidate_cache(f);
 525                 }
 526                 if (ret < 0 && errno == EACCES && geteuid()) {
 527                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 528                                 log_err("fio: only root may flush block "
 529                                         "devices. Cache flush bypassed!\n");
 530                         }
 531                 }
 532                 if (ret < 0)
 533                         errval = errno;
 534         } else if (f->filetype == FIO_TYPE_CHAR ||
 535                    f->filetype == FIO_TYPE_PIPE) {
 536                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 537         }
 538
 539         /*
 540          * Cache flushing isn't a fatal condition, and we know it will
 541          * happen on some platforms where we don't have the proper
 542          * function to flush eg block device caches. So just warn and
 543          * continue on our way.
 544          */
 545         if (errval)
 546                 log_info("fio: cache invalidation of %s failed: %s\n",
 547                          f->file_name, strerror(errval));
 548
 549         return 0;
 550
 551 }
 552
 553 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 554 {
 555         if (!fio_file_open(f))
 556                 return 0;
 557
 558         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 559 }
 560
 561 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 562 {
 563         int ret = 0;
 564
 565         dprint(FD_FILE, "fd close %s\n", f->file_name);
 566
 567         remove_file_hash(f);
 568
 569         if (close(f->fd) < 0)
 570                 ret = errno;
 571
 572         f->fd = -1;
 573
 574         if (f->shadow_fd != -1) {
 575                 close(f->shadow_fd);
 576                 f->shadow_fd = -1;
 577         }
 578
 579         f->engine_pos = 0;
 580         return ret;
 581 }
 582
 583 int file_lookup_open(struct fio_file *f, int flags)
 584 {
 585         struct fio_file *__f;
 586         int from_hash;
 587
 588         __f = lookup_file_hash(f->file_name);
 589         if (__f) {
 590                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 591                 f->lock = __f->lock;
 592                 from_hash = 1;
 593         } else {
 594                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 595                 from_hash = 0;
 596         }
 597
 598 #ifdef WIN32
 599         flags |= _O_BINARY;
 600 #endif
 601
 602         f->fd = open(f->file_name, flags, 0600);
 603         return from_hash;
 604 }
 605
 606 static int file_close_shadow_fds(struct thread_data *td)
 607 {
 608         struct fio_file *f;
 609         int num_closed = 0;
 610         unsigned int i;
 611
 612         for_each_file(td, f, i) {
 613                 if (f->shadow_fd == -1)
 614                         continue;
 615
 616                 close(f->shadow_fd);
 617                 f->shadow_fd = -1;
 618                 num_closed++;
 619         }
 620
 621         return num_closed;
 622 }
 623
 624 int generic_open_file(struct thread_data *td, struct fio_file *f)
 625 {
 626         int is_std = 0;
 627         int flags = 0;
 628         int from_hash = 0;
 629
 630         dprint(FD_FILE, "fd open %s\n", f->file_name);
 631
 632         if (!strcmp(f->file_name, "-")) {
 633                 if (td_rw(td)) {
 634                         log_err("fio: can't read/write to stdin/out\n");
 635                         return 1;
 636                 }
 637                 is_std = 1;
 638
 639                 /*
 640                  * move output logging to stderr, if we are writing to stdout
 641                  */
 642                 if (td_write(td))
 643                         f_out = stderr;
 644         }
 645
 646         if (td_trim(td))
 647                 goto skip_flags;
 648         if (td->o.odirect)
 649                 flags |= OS_O_DIRECT;
 650         if (td->o.oatomic) {
 651                 if (!FIO_O_ATOMIC) {
 652                         td_verror(td, EINVAL, "OS does not support atomic IO");
 653                         return 1;
 654                 }
 655                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 656         }
 657         flags |= td->o.sync_io;
 658         if (td->o.create_on_open && td->o.allow_create)
 659                 flags |= O_CREAT;
 660 skip_flags:
 661         if (f->filetype != FIO_TYPE_FILE)
 662                 flags |= FIO_O_NOATIME;
 663
 664 open_again:
 665         if (td_write(td)) {
 666                 if (!read_only)
 667                         flags |= O_RDWR;
 668
 669                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 670                         flags |= O_CREAT;
 671
 672                 if (is_std)
 673                         f->fd = dup(STDOUT_FILENO);
 674                 else
 675                         from_hash = file_lookup_open(f, flags);
 676         } else if (td_read(td)) {
 677                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 678                         flags |= O_RDWR;
 679                 else
 680                         flags |= O_RDONLY;
 681
 682                 if (is_std)
 683                         f->fd = dup(STDIN_FILENO);
 684                 else
 685                         from_hash = file_lookup_open(f, flags);
 686         } else if (td_trim(td)) {
 687                 assert(!td_rw(td)); /* should have matched above */
 688                 if (!read_only)
 689                         flags |= O_RDWR;
 690                 from_hash = file_lookup_open(f, flags);
 691         }
 692
 693         if (f->fd == -1) {
 694                 char buf[FIO_VERROR_SIZE];
 695                 int __e = errno;
 696
 697                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 698                         flags &= ~FIO_O_NOATIME;
 699                         goto open_again;
 700                 }
 701                 if (__e == EMFILE && file_close_shadow_fds(td))
 702                         goto open_again;
 703
 704                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 705
 706                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 707                         log_err("fio: looks like your file system does not " \
 708                                 "support direct=1/buffered=0\n");
 709                 }
 710
 711                 td_verror(td, __e, buf);
 712                 return 1;
 713         }
 714
 715         if (!from_hash && f->fd != -1) {
 716                 if (add_file_hash(f)) {
 717                         int fio_unused ret;
 718
 719                         /*
 720                          * Stash away descriptor for later close. This is to
 721                          * work-around a "feature" on Linux, where a close of
 722                          * an fd that has been opened for write will trigger
 723                          * udev to call blkid to check partitions, fs id, etc.
 724                          * That pollutes the device cache, which can slow down
 725                          * unbuffered accesses.
 726                          */
 727                         if (f->shadow_fd == -1)
 728                                 f->shadow_fd = f->fd;
 729                         else {
 730                                 /*
 731                                  * OK to ignore, we haven't done anything
 732                                  * with it
 733                                  */
 734                                 ret = generic_close_file(td, f);
 735                         }
 736                         goto open_again;
 737                 }
 738         }
 739
 740         return 0;
 741 }
 742
 743 /*
 744  * This function i.e. get_file_size() is the default .get_file_size
 745  * implementation of majority of I/O engines.
 746  */
 747 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 748 {
 749         return get_file_size(td, f);
 750 }
 751
 752 /*
 753  * open/close all files, so that ->real_file_size gets set
 754  */
 755 static int get_file_sizes(struct thread_data *td)
 756 {
 757         struct fio_file *f;
 758         unsigned int i;
 759         int err = 0;
 760
 761         for_each_file(td, f, i) {
 762                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 763                                                                 f->file_name);
 764
 765                 if (td_io_get_file_size(td, f)) {
 766                         if (td->error != ENOENT) {
 767                                 log_err("%s\n", td->verror);
 768                                 err = 1;
 769                                 break;
 770                         }
 771                         clear_error(td);
 772                 }
 773
 774                 /*
 775                  * There are corner cases where we end up with -1 for
 776                  * ->real_file_size due to unsupported file type, etc.
 777                  * We then just set to size option value divided by number
 778                  * of files, similar to the way file ->io_size is set.
 779                  * stat(2) failure doesn't set ->real_file_size to -1.
 780                  */
 781                 if (f->real_file_size == -1ULL && td->o.size)
 782                         f->real_file_size = td->o.size / td->o.nr_files;
 783         }
 784
 785         return err;
 786 }
 787
 788 struct fio_mount {
 789         struct flist_head list;
 790         const char *base;
 791         char __base[256];
 792         unsigned int key;
 793 };
 794
 795 /*
 796  * Get free number of bytes for each file on each unique mount.
 797  */
 798 static unsigned long long get_fs_free_counts(struct thread_data *td)
 799 {
 800         struct flist_head *n, *tmp;
 801         unsigned long long ret = 0;
 802         struct fio_mount *fm;
 803         FLIST_HEAD(list);
 804         struct fio_file *f;
 805         unsigned int i;
 806
 807         for_each_file(td, f, i) {
 808                 struct stat sb;
 809                 char buf[256];
 810
 811                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 812                         if (f->real_file_size != -1ULL)
 813                                 ret += f->real_file_size;
 814                         continue;
 815                 } else if (f->filetype != FIO_TYPE_FILE)
 816                         continue;
 817
 818                 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
 819
 820                 if (stat(buf, &sb) < 0) {
 821                         if (errno != ENOENT)
 822                                 break;
 823                         strcpy(buf, ".");
 824                         if (stat(buf, &sb) < 0)
 825                                 break;
 826                 }
 827
 828                 fm = NULL;
 829                 flist_for_each(n, &list) {
 830                         fm = flist_entry(n, struct fio_mount, list);
 831                         if (fm->key == sb.st_dev)
 832                                 break;
 833
 834                         fm = NULL;
 835                 }
 836
 837                 if (fm)
 838                         continue;
 839
 840                 fm = calloc(1, sizeof(*fm));
 841                 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
 842                 fm->base = basename(fm->__base);
 843                 fm->key = sb.st_dev;
 844                 flist_add(&fm->list, &list);
 845         }
 846
 847         flist_for_each_safe(n, tmp, &list) {
 848                 unsigned long long sz;
 849
 850                 fm = flist_entry(n, struct fio_mount, list);
 851                 flist_del(&fm->list);
 852
 853                 sz = get_fs_free_size(fm->base);
 854                 if (sz && sz != -1ULL)
 855                         ret += sz;
 856
 857                 free(fm);
 858         }
 859
 860         return ret;
 861 }
 862
 863 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 864 {
 865         bool align = false;
 866         struct thread_options *o = &td->o;
 867         unsigned long long align_bs;
 868         unsigned long long offset;
 869         unsigned long long increment;
 870
 871         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 872                 return f->real_file_size;
 873
 874         if (o->offset_increment_percent) {
 875                 assert(!o->offset_increment);
 876                 increment = o->offset_increment_percent * f->real_file_size / 100;
 877                 align = true;
 878         } else
 879                 increment = o->offset_increment;
 880
 881         if (o->start_offset_percent > 0) {
 882                 /* calculate the raw offset */
 883                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 884                         (td->subjob_number * increment);
 885
 886                 align = true;
 887         } else {
 888                 /* start_offset_percent not set */
 889                 offset = o->start_offset +
 890                                 td->subjob_number * increment;
 891         }
 892
 893         if (align) {
 894                 /*
 895                  * if offset_align is provided, use it
 896                  */
 897                 if (fio_option_is_set(o, start_offset_align)) {
 898                         align_bs = o->start_offset_align;
 899                 } else {
 900                         /* else take the minimum block size */
 901                         align_bs = td_min_bs(td);
 902                 }
 903
 904                 /*
 905                  * block align the offset at the next available boundary at
 906                  * ceiling(offset / align_bs) * align_bs
 907                  */
 908                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 909         }
 910
 911         return offset;
 912 }
 913
 914 /*
 915  * Find longest path component that exists and return its length
 916  */
 917 int longest_existing_path(char *path) {
 918         char buf[PATH_MAX];
 919         bool done;
 920         char *buf_pos;
 921         int offset;
 922 #ifdef WIN32
 923         DWORD dwAttr;
 924 #else
 925         struct stat sb;
 926 #endif
 927
 928         sprintf(buf, "%s", path);
 929         done = false;
 930         while (!done) {
 931                 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
 932                 if (!buf_pos) {
 933                         done = true;
 934                         offset = 0;
 935                         break;
 936                 }
 937
 938                 *(buf_pos + 1) = '\0';
 939
 940 #ifdef WIN32
 941                 dwAttr = GetFileAttributesA(buf);
 942                 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
 943                         done = true;
 944                 }
 945 #else
 946                 if (stat(buf, &sb) == 0)
 947                         done = true;
 948 #endif
 949                 if (done)
 950                         offset = buf_pos - buf;
 951                 else
 952                         *buf_pos = '\0';
 953         }
 954
 955         return offset;
 956 }
 957
 958 static bool create_work_dirs(struct thread_data *td, const char *fname)
 959 {
 960         char path[PATH_MAX];
 961         char *start, *end;
 962         int offset;
 963
 964         snprintf(path, PATH_MAX, "%s", fname);
 965         start = path;
 966
 967         offset = longest_existing_path(path);
 968         end = start + offset;
 969         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
 970                 if (end == start) {
 971                         end++;
 972                         continue;
 973                 }
 974                 *end = '\0';
 975                 errno = 0;
 976                 if (fio_mkdir(path, 0700) && errno != EEXIST) {
 977                         log_err("fio: failed to create dir (%s): %s\n",
 978                                 start, strerror(errno));
 979                         return false;
 980                 }
 981                 *end = FIO_OS_PATH_SEPARATOR;
 982                 end++;
 983         }
 984         td->flags |= TD_F_DIRS_CREATED;
 985         return true;
 986 }
 987
 988 /*
 989  * Open the files and setup files sizes, creating files if necessary.
 990  */
 991 int setup_files(struct thread_data *td)
 992 {
 993         unsigned long long total_size, extend_size;
 994         struct thread_options *o = &td->o;
 995         struct fio_file *f;
 996         unsigned int i, nr_fs_extra = 0;
 997         int err = 0, need_extend;
 998         int old_state;
 999         const unsigned long long bs = td_min_bs(td);
1000         uint64_t fs = 0;
1001
1002         dprint(FD_FILE, "setup files\n");
1003
1004         old_state = td_bump_runstate(td, TD_SETTING_UP);
1005
1006         for_each_file(td, f, i) {
1007                 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1008                     strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1009                     !(td->flags & TD_F_DIRS_CREATED) &&
1010                     !create_work_dirs(td, f->file_name))
1011                         goto err_out;
1012         }
1013
1014         /*
1015          * Find out physical size of files or devices for this thread,
1016          * before we determine I/O size and range of our targets.
1017          * If ioengine defines a setup() method, it's responsible for
1018          * opening the files and setting f->real_file_size to indicate
1019          * the valid range for that file.
1020          */
1021         if (td->io_ops->setup)
1022                 err = td->io_ops->setup(td);
1023         else
1024                 err = get_file_sizes(td);
1025
1026         if (err)
1027                 goto err_out;
1028
1029         if (o->read_iolog_file)
1030                 goto done;
1031
1032         /*
1033          * check sizes. if the files/devices do not exist and the size
1034          * isn't passed to fio, abort.
1035          */
1036         total_size = 0;
1037         for_each_file(td, f, i) {
1038                 f->fileno = i;
1039                 if (f->real_file_size == -1ULL)
1040                         total_size = -1ULL;
1041                 else
1042                         total_size += f->real_file_size;
1043         }
1044
1045         if (o->fill_device)
1046                 td->fill_device_size = get_fs_free_counts(td);
1047
1048         /*
1049          * device/file sizes are zero and no size given, punt
1050          */
1051         if ((!total_size || total_size == -1ULL) && !o->size &&
1052             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1053             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1054                 log_err("%s: you need to specify size=\n", o->name);
1055                 td_verror(td, EINVAL, "total_file_size");
1056                 goto err_out;
1057         }
1058
1059         /*
1060          * Calculate per-file size and potential extra size for the
1061          * first files, if needed (i.e. if we don't have a fixed size).
1062          */
1063         if (!o->file_size_low && o->nr_files) {
1064                 uint64_t all_fs;
1065
1066                 fs = o->size / o->nr_files;
1067                 all_fs = fs * o->nr_files;
1068
1069                 if (all_fs < o->size)
1070                         nr_fs_extra = (o->size - all_fs) / bs;
1071         }
1072
1073         /*
1074          * now file sizes are known, so we can set ->io_size. if size= is
1075          * not given, ->io_size is just equal to ->real_file_size. if size
1076          * is given, ->io_size is size / nr_files.
1077          */
1078         extend_size = total_size = 0;
1079         need_extend = 0;
1080         for_each_file(td, f, i) {
1081                 f->file_offset = get_start_offset(td, f);
1082
1083                 /*
1084                  * Update ->io_size depending on options specified.
1085                  * ->file_size_low being 0 means filesize option isn't set.
1086                  * Non zero ->file_size_low equals ->file_size_high means
1087                  * filesize option is set in a fixed size format.
1088                  * Non zero ->file_size_low not equals ->file_size_high means
1089                  * filesize option is set in a range format.
1090                  */
1091                 if (!o->file_size_low) {
1092                         /*
1093                          * no file size or range given, file size is equal to
1094                          * total size divided by number of files. If the size
1095                          * doesn't divide nicely with the min blocksize,
1096                          * make the first files bigger.
1097                          */
1098                         f->io_size = fs;
1099                         if (nr_fs_extra) {
1100                                 nr_fs_extra--;
1101                                 f->io_size += bs;
1102                         }
1103
1104                         /*
1105                          * We normally don't come here for regular files, but
1106                          * if the result is 0 for a regular file, set it to the
1107                          * real file size. This could be size of the existing
1108                          * one if it already exists, but otherwise will be set
1109                          * to 0. A new file won't be created because
1110                          * ->io_size + ->file_offset equals ->real_file_size.
1111                          */
1112                         if (!f->io_size) {
1113                                 if (f->file_offset > f->real_file_size)
1114                                         goto err_offset;
1115                                 f->io_size = f->real_file_size - f->file_offset;
1116                                 if (!f->io_size)
1117                                         log_info("fio: file %s may be ignored\n",
1118                                                 f->file_name);
1119                         }
1120                 } else if (f->real_file_size < o->file_size_low ||
1121                            f->real_file_size > o->file_size_high) {
1122                         if (f->file_offset > o->file_size_low)
1123                                 goto err_offset;
1124                         /*
1125                          * file size given. if it's fixed, use that. if it's a
1126                          * range, generate a random size in-between.
1127                          */
1128                         if (o->file_size_low == o->file_size_high)
1129                                 f->io_size = o->file_size_low - f->file_offset;
1130                         else {
1131                                 f->io_size = get_rand_file_size(td)
1132                                                 - f->file_offset;
1133                         }
1134                 } else
1135                         f->io_size = f->real_file_size - f->file_offset;
1136
1137                 if (f->io_size == -1ULL)
1138                         total_size = -1ULL;
1139                 else {
1140                         uint64_t io_size;
1141
1142                         if (o->size_percent && o->size_percent != 100) {
1143                                 uint64_t file_size;
1144
1145                                 file_size = f->io_size + f->file_offset;
1146                                 f->io_size = (file_size *
1147                                               o->size_percent) / 100;
1148                                 if (f->io_size > (file_size - f->file_offset))
1149                                         f->io_size = file_size - f->file_offset;
1150
1151                                 f->io_size -= (f->io_size % td_min_bs(td));
1152                         }
1153
1154                         io_size = f->io_size;
1155                         if (o->io_size_percent && o->io_size_percent != 100) {
1156                                 io_size *= o->io_size_percent;
1157                                 io_size /= 100;
1158                         }
1159
1160                         total_size += io_size;
1161                 }
1162
1163                 if (f->filetype == FIO_TYPE_FILE &&
1164                     (f->io_size + f->file_offset) > f->real_file_size) {
1165                         if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1166                             !o->create_on_open) {
1167                                 need_extend++;
1168                                 extend_size += (f->io_size + f->file_offset);
1169                                 fio_file_set_extend(f);
1170                         } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1171                                    (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1172                                     td_ioengine_flagged(td, FIO_FAKEIO)))
1173                                 f->real_file_size = f->io_size + f->file_offset;
1174                 }
1175         }
1176
1177         if (td->o.block_error_hist) {
1178                 int len;
1179
1180                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1181                 f = td->files[0];
1182                 len = f->io_size / td->o.bs[DDIR_TRIM];
1183                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1184                         log_err("fio: cannot calculate block histogram with "
1185                                 "%d trim blocks, maximum %d\n",
1186                                 len, MAX_NR_BLOCK_INFOS);
1187                         td_verror(td, EINVAL, "block_error_hist");
1188                         goto err_out;
1189                 }
1190
1191                 td->ts.nr_block_infos = len;
1192                 for (i = 0; i < len; i++)
1193                         td->ts.block_infos[i] =
1194                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1195         } else
1196                 td->ts.nr_block_infos = 0;
1197
1198         if (!o->size || (total_size && o->size > total_size))
1199                 o->size = total_size;
1200
1201         if (o->size < td_min_bs(td)) {
1202                 log_err("fio: blocksize is larger than data set range\n");
1203                 goto err_out;
1204         }
1205
1206         /*
1207          * See if we need to extend some files, typically needed when our
1208          * target regular files don't exist yet, but our jobs require them
1209          * initially due to read I/Os.
1210          */
1211         if (need_extend) {
1212                 temp_stall_ts = 1;
1213                 if (output_format & FIO_OUTPUT_NORMAL) {
1214                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1215                                  o->name,
1216                                  need_extend > 1 ? "s" : "",
1217                                  need_extend,
1218                                  need_extend > 1 ? "s" : "",
1219                                  need_extend > 1 ? "total " : "",
1220                                  extend_size >> 20);
1221                 }
1222
1223                 for_each_file(td, f, i) {
1224                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1225
1226                         if (!fio_file_extend(f))
1227                                 continue;
1228
1229                         assert(f->filetype == FIO_TYPE_FILE);
1230                         fio_file_clear_extend(f);
1231                         if (!o->fill_device) {
1232                                 old_len = f->real_file_size;
1233                                 extend_len = f->io_size + f->file_offset -
1234                                                 old_len;
1235                         }
1236                         f->real_file_size = (f->io_size + f->file_offset);
1237                         err = extend_file(td, f);
1238                         if (err)
1239                                 break;
1240
1241                         err = __file_invalidate_cache(td, f, old_len,
1242                                                                 extend_len);
1243
1244                         /*
1245                          * Shut up static checker
1246                          */
1247                         if (f->fd != -1)
1248                                 close(f->fd);
1249
1250                         f->fd = -1;
1251                         if (err)
1252                                 break;
1253                 }
1254                 temp_stall_ts = 0;
1255         }
1256
1257         if (err)
1258                 goto err_out;
1259
1260         /*
1261          * iolog already set the total io size, if we read back
1262          * stored entries.
1263          */
1264         if (!o->read_iolog_file) {
1265                 if (o->io_size)
1266                         td->total_io_size = o->io_size * o->loops;
1267                 else
1268                         td->total_io_size = o->size * o->loops;
1269         }
1270
1271 done:
1272         if (o->create_only)
1273                 td->done = 1;
1274
1275         td_restore_runstate(td, old_state);
1276
1277         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1278                 err = zbd_setup_files(td);
1279                 if (err)
1280                         goto err_out;
1281         }
1282         return 0;
1283
1284 err_offset:
1285         log_err("%s: you need to specify valid offset=\n", o->name);
1286 err_out:
1287         td_restore_runstate(td, old_state);
1288         return 1;
1289 }
1290
1291 bool pre_read_files(struct thread_data *td)
1292 {
1293         struct fio_file *f;
1294         unsigned int i;
1295
1296         dprint(FD_FILE, "pre_read files\n");
1297
1298         for_each_file(td, f, i) {
1299                 if (!pre_read_file(td, f))
1300                         return false;
1301         }
1302
1303         return true;
1304 }
1305
1306 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1307 {
1308         unsigned int range_size, seed;
1309         uint64_t nranges;
1310         uint64_t fsize;
1311
1312         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1313         fsize = min(f->real_file_size, f->io_size);
1314
1315         nranges = (fsize + range_size - 1ULL) / range_size;
1316
1317         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1318         if (!td->o.rand_repeatable)
1319                 seed = td->rand_seeds[4];
1320
1321         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1322                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1323         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1324                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1325         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1326                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1327 }
1328
1329 static bool init_rand_distribution(struct thread_data *td)
1330 {
1331         struct fio_file *f;
1332         unsigned int i;
1333         int state;
1334
1335         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1336             td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1337             td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1338                 return false;
1339
1340         state = td_bump_runstate(td, TD_SETTING_UP);
1341
1342         for_each_file(td, f, i)
1343                 __init_rand_distribution(td, f);
1344
1345         td_restore_runstate(td, state);
1346         return true;
1347 }
1348
1349 /*
1350  * Check if the number of blocks exceeds the randomness capability of
1351  * the selected generator. Tausworthe is 32-bit, the others are fullly
1352  * 64-bit capable.
1353  */
1354 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1355                                  uint64_t blocks)
1356 {
1357         if (blocks <= FRAND32_MAX)
1358                 return 0;
1359         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1360                 return 0;
1361
1362         /*
1363          * If the user hasn't specified a random generator, switch
1364          * to tausworthe64 with informational warning. If the user did
1365          * specify one, just warn.
1366          */
1367         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1368                         f->file_name);
1369
1370         if (!fio_option_is_set(&td->o, random_generator)) {
1371                 log_info("fio: Switching to tausworthe64. Use the "
1372                          "random_generator= option to get rid of this "
1373                          "warning.\n");
1374                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1375                 return 0;
1376         }
1377
1378         /*
1379          * Just make this information to avoid breaking scripts.
1380          */
1381         log_info("fio: Use the random_generator= option to switch to lfsr or "
1382                          "tausworthe64.\n");
1383         return 0;
1384 }
1385
1386 bool init_random_map(struct thread_data *td)
1387 {
1388         unsigned long long blocks;
1389         struct fio_file *f;
1390         unsigned int i;
1391
1392         if (init_rand_distribution(td))
1393                 return true;
1394         if (!td_random(td))
1395                 return true;
1396
1397         for_each_file(td, f, i) {
1398                 uint64_t fsize = min(f->real_file_size, f->io_size);
1399
1400                 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1401                         fsize = td->o.zone_range;
1402
1403                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1404
1405                 if (check_rand_gen_limits(td, f, blocks))
1406                         return false;
1407
1408                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1409                         uint64_t seed;
1410
1411                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1412
1413                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1414                                 fio_file_set_lfsr(f);
1415                                 continue;
1416                         } else {
1417                                 log_err("fio: failed initializing LFSR\n");
1418                                 return false;
1419                         }
1420                 } else if (!td->o.norandommap) {
1421                         f->io_axmap = axmap_new(blocks);
1422                         if (f->io_axmap) {
1423                                 fio_file_set_axmap(f);
1424                                 continue;
1425                         }
1426                 } else if (td->o.norandommap)
1427                         continue;
1428
1429                 if (!td->o.softrandommap) {
1430                         log_err("fio: failed allocating random map. If running"
1431                                 " a large number of jobs, try the 'norandommap'"
1432                                 " option or set 'softrandommap'. Or give"
1433                                 " a larger --alloc-size to fio.\n");
1434                         return false;
1435                 }
1436
1437                 log_info("fio: file %s failed allocating random map. Running "
1438                          "job without.\n", f->file_name);
1439         }
1440
1441         return true;
1442 }
1443
1444 void close_files(struct thread_data *td)
1445 {
1446         struct fio_file *f;
1447         unsigned int i;
1448
1449         for_each_file(td, f, i) {
1450                 if (fio_file_open(f))
1451                         td_io_close_file(td, f);
1452         }
1453 }
1454
1455 void fio_file_free(struct fio_file *f)
1456 {
1457         if (fio_file_axmap(f))
1458                 axmap_free(f->io_axmap);
1459         if (!fio_file_smalloc(f)) {
1460                 free(f->file_name);
1461                 free(f);
1462         } else {
1463                 sfree(f->file_name);
1464                 sfree(f);
1465         }
1466 }
1467
1468 void close_and_free_files(struct thread_data *td)
1469 {
1470         struct fio_file *f;
1471         unsigned int i;
1472
1473         dprint(FD_FILE, "close files\n");
1474
1475         for_each_file(td, f, i) {
1476                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1477                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1478                         td_io_unlink_file(td, f);
1479                 }
1480
1481                 if (fio_file_open(f))
1482                         td_io_close_file(td, f);
1483
1484                 remove_file_hash(f);
1485
1486                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1487                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1488                         td_io_unlink_file(td, f);
1489                 }
1490
1491                 zbd_close_file(f);
1492                 fio_file_free(f);
1493         }
1494
1495         td->o.filename = NULL;
1496         free(td->files);
1497         free(td->file_locks);
1498         td->files_index = 0;
1499         td->files = NULL;
1500         td->file_locks = NULL;
1501         td->o.file_lock_mode = FILE_LOCK_NONE;
1502         td->o.nr_files = 0;
1503 }
1504
1505 static void get_file_type(struct fio_file *f)
1506 {
1507         struct stat sb;
1508
1509         if (!strcmp(f->file_name, "-"))
1510                 f->filetype = FIO_TYPE_PIPE;
1511         else
1512                 f->filetype = FIO_TYPE_FILE;
1513
1514 #ifdef WIN32
1515         /* \\.\ is the device namespace in Windows, where every file is
1516          * a block device */
1517         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1518                 f->filetype = FIO_TYPE_BLOCK;
1519 #endif
1520
1521         if (!stat(f->file_name, &sb)) {
1522                 if (S_ISBLK(sb.st_mode))
1523                         f->filetype = FIO_TYPE_BLOCK;
1524                 else if (S_ISCHR(sb.st_mode))
1525                         f->filetype = FIO_TYPE_CHAR;
1526                 else if (S_ISFIFO(sb.st_mode))
1527                         f->filetype = FIO_TYPE_PIPE;
1528         }
1529 }
1530
1531 static bool __is_already_allocated(const char *fname, bool set)
1532 {
1533         struct flist_head *entry;
1534         bool ret;
1535
1536         ret = file_bloom_exists(fname, set);
1537         if (!ret)
1538                 return ret;
1539
1540         flist_for_each(entry, &filename_list) {
1541                 struct file_name *fn;
1542
1543                 fn = flist_entry(entry, struct file_name, list);
1544
1545                 if (!strcmp(fn->filename, fname))
1546                         return true;
1547         }
1548
1549         return false;
1550 }
1551
1552 static bool is_already_allocated(const char *fname)
1553 {
1554         bool ret;
1555
1556         fio_file_hash_lock();
1557         ret = __is_already_allocated(fname, false);
1558         fio_file_hash_unlock();
1559
1560         return ret;
1561 }
1562
1563 static void set_already_allocated(const char *fname)
1564 {
1565         struct file_name *fn;
1566
1567         fn = malloc(sizeof(struct file_name));
1568         fn->filename = strdup(fname);
1569
1570         fio_file_hash_lock();
1571         if (!__is_already_allocated(fname, true)) {
1572                 flist_add_tail(&fn->list, &filename_list);
1573                 fn = NULL;
1574         }
1575         fio_file_hash_unlock();
1576
1577         if (fn) {
1578                 free(fn->filename);
1579                 free(fn);
1580         }
1581 }
1582
1583 static void free_already_allocated(void)
1584 {
1585         struct flist_head *entry, *tmp;
1586         struct file_name *fn;
1587
1588         if (flist_empty(&filename_list))
1589                 return;
1590
1591         fio_file_hash_lock();
1592         flist_for_each_safe(entry, tmp, &filename_list) {
1593                 fn = flist_entry(entry, struct file_name, list);
1594                 free(fn->filename);
1595                 flist_del(&fn->list);
1596                 free(fn);
1597         }
1598
1599         fio_file_hash_unlock();
1600 }
1601
1602 static struct fio_file *alloc_new_file(struct thread_data *td)
1603 {
1604         struct fio_file *f;
1605
1606         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1607                 f = calloc(1, sizeof(*f));
1608         else
1609                 f = scalloc(1, sizeof(*f));
1610         if (!f) {
1611                 assert(0);
1612                 return NULL;
1613         }
1614
1615         f->fd = -1;
1616         f->shadow_fd = -1;
1617         fio_file_reset(td, f);
1618         if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1619                 fio_file_set_smalloc(f);
1620         return f;
1621 }
1622
1623 bool exists_and_not_regfile(const char *filename)
1624 {
1625         struct stat sb;
1626
1627         if (lstat(filename, &sb) == -1)
1628                 return false;
1629
1630 #ifndef WIN32 /* NOT Windows */
1631         if (S_ISREG(sb.st_mode))
1632                 return false;
1633 #else
1634         /* \\.\ is the device namespace in Windows, where every file
1635          * is a device node */
1636         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1637                 return false;
1638 #endif
1639
1640         return true;
1641 }
1642
1643 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1644 {
1645         int cur_files = td->files_index;
1646         char file_name[PATH_MAX];
1647         struct fio_file *f;
1648         int len = 0;
1649
1650         dprint(FD_FILE, "add file %s\n", fname);
1651
1652         if (td->o.directory)
1653                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1654                                         td->o.unique_filename);
1655
1656         sprintf(file_name + len, "%s", fname);
1657
1658         /* clean cloned siblings using existing files */
1659         if (numjob && is_already_allocated(file_name) &&
1660             !exists_and_not_regfile(fname))
1661                 return 0;
1662
1663         f = alloc_new_file(td);
1664
1665         if (td->files_size <= td->files_index) {
1666                 unsigned int new_size = td->o.nr_files + 1;
1667
1668                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1669
1670                 td->files = realloc(td->files, new_size * sizeof(f));
1671                 if (td->files == NULL) {
1672                         log_err("fio: realloc OOM\n");
1673                         assert(0);
1674                 }
1675                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1676                         td->file_locks = realloc(td->file_locks, new_size);
1677                         if (!td->file_locks) {
1678                                 log_err("fio: realloc OOM\n");
1679                                 assert(0);
1680                         }
1681                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1682                 }
1683                 td->files_size = new_size;
1684         }
1685         td->files[cur_files] = f;
1686         f->fileno = cur_files;
1687
1688         /*
1689          * init function, io engine may not be loaded yet
1690          */
1691         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1692                 f->real_file_size = -1ULL;
1693
1694         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1695                 f->file_name = strdup(file_name);
1696         else
1697                 f->file_name = smalloc_strdup(file_name);
1698
1699         /* can't handle smalloc failure from here */
1700         assert(f->file_name);
1701
1702         get_file_type(f);
1703
1704         switch (td->o.file_lock_mode) {
1705         case FILE_LOCK_NONE:
1706                 break;
1707         case FILE_LOCK_READWRITE:
1708                 f->rwlock = fio_rwlock_init();
1709                 break;
1710         case FILE_LOCK_EXCLUSIVE:
1711                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1712                 break;
1713         default:
1714                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1715                 assert(0);
1716         }
1717
1718         td->files_index++;
1719
1720         if (td->o.numjobs > 1)
1721                 set_already_allocated(file_name);
1722
1723         if (inc)
1724                 td->o.nr_files++;
1725
1726         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1727                                                         cur_files);
1728
1729         return cur_files;
1730 }
1731
1732 int add_file_exclusive(struct thread_data *td, const char *fname)
1733 {
1734         struct fio_file *f;
1735         unsigned int i;
1736
1737         for_each_file(td, f, i) {
1738                 if (!strcmp(f->file_name, fname))
1739                         return i;
1740         }
1741
1742         return add_file(td, fname, 0, 1);
1743 }
1744
1745 void get_file(struct fio_file *f)
1746 {
1747         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1748         assert(fio_file_open(f));
1749         f->references++;
1750 }
1751
1752 int put_file(struct thread_data *td, struct fio_file *f)
1753 {
1754         int f_ret = 0, ret = 0;
1755
1756         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1757
1758         if (!fio_file_open(f)) {
1759                 assert(f->fd == -1);
1760                 return 0;
1761         }
1762
1763         assert(f->references);
1764         if (--f->references)
1765                 return 0;
1766
1767         disk_util_dec(f->du);
1768
1769         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1770                 unlock_file_all(td, f);
1771
1772         if (should_fsync(td) && td->o.fsync_on_close) {
1773                 f_ret = fsync(f->fd);
1774                 if (f_ret < 0)
1775                         f_ret = errno;
1776         }
1777
1778         if (td->io_ops->close_file)
1779                 ret = td->io_ops->close_file(td, f);
1780
1781         if (!ret)
1782                 ret = f_ret;
1783
1784         td->nr_open_files--;
1785         fio_file_clear_closing(f);
1786         fio_file_clear_open(f);
1787         assert(f->fd == -1);
1788         return ret;
1789 }
1790
1791 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1792 {
1793         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1794                 return;
1795
1796         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1797                 if (ddir == DDIR_READ)
1798                         fio_rwlock_read(f->rwlock);
1799                 else
1800                         fio_rwlock_write(f->rwlock);
1801         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1802                 fio_sem_down(f->lock);
1803
1804         td->file_locks[f->fileno] = td->o.file_lock_mode;
1805 }
1806
1807 void unlock_file(struct thread_data *td, struct fio_file *f)
1808 {
1809         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1810                 return;
1811
1812         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1813                 fio_rwlock_unlock(f->rwlock);
1814         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1815                 fio_sem_up(f->lock);
1816
1817         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1818 }
1819
1820 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1821 {
1822         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1823                 return;
1824         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1825                 unlock_file(td, f);
1826 }
1827
1828 static bool recurse_dir(struct thread_data *td, const char *dirname)
1829 {
1830         struct dirent *dir;
1831         bool ret = false;
1832         DIR *D;
1833
1834         D = opendir(dirname);
1835         if (!D) {
1836                 char buf[FIO_VERROR_SIZE];
1837
1838                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1839                 td_verror(td, errno, buf);
1840                 return true;
1841         }
1842
1843         while ((dir = readdir(D)) != NULL) {
1844                 char full_path[PATH_MAX];
1845                 struct stat sb;
1846
1847                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1848                         continue;
1849
1850                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1851
1852                 if (lstat(full_path, &sb) == -1) {
1853                         if (errno != ENOENT) {
1854                                 td_verror(td, errno, "stat");
1855                                 ret = true;
1856                                 break;
1857                         }
1858                 }
1859
1860                 if (S_ISREG(sb.st_mode)) {
1861                         add_file(td, full_path, 0, 1);
1862                         continue;
1863                 }
1864                 if (!S_ISDIR(sb.st_mode))
1865                         continue;
1866
1867                 ret = recurse_dir(td, full_path);
1868                 if (ret)
1869                         break;
1870         }
1871
1872         closedir(D);
1873         return ret;
1874 }
1875
1876 int add_dir_files(struct thread_data *td, const char *path)
1877 {
1878         int ret = recurse_dir(td, path);
1879
1880         if (!ret)
1881                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1882
1883         return ret;
1884 }
1885
1886 void dup_files(struct thread_data *td, struct thread_data *org)
1887 {
1888         struct fio_file *f;
1889         unsigned int i;
1890
1891         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1892
1893         if (!org->files)
1894                 return;
1895
1896         td->files = malloc(org->files_index * sizeof(f));
1897
1898         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1899                 td->file_locks = malloc(org->files_index);
1900
1901         for_each_file(org, f, i) {
1902                 struct fio_file *__f;
1903
1904                 __f = alloc_new_file(td);
1905
1906                 if (f->file_name) {
1907                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1908                                 __f->file_name = strdup(f->file_name);
1909                         else
1910                                 __f->file_name = smalloc_strdup(f->file_name);
1911
1912                         /* can't handle smalloc failure from here */
1913                         assert(__f->file_name);
1914                         __f->filetype = f->filetype;
1915                 }
1916
1917                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1918                         __f->lock = f->lock;
1919                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1920                         __f->rwlock = f->rwlock;
1921
1922                 td->files[i] = __f;
1923         }
1924 }
1925
1926 /*
1927  * Returns the index that matches the filename, or -1 if not there
1928  */
1929 int get_fileno(struct thread_data *td, const char *fname)
1930 {
1931         struct fio_file *f;
1932         unsigned int i;
1933
1934         for_each_file(td, f, i)
1935                 if (!strcmp(f->file_name, fname))
1936                         return i;
1937
1938         return -1;
1939 }
1940
1941 /*
1942  * For log usage, where we add/open/close files automatically
1943  */
1944 void free_release_files(struct thread_data *td)
1945 {
1946         close_files(td);
1947         td->o.nr_files = 0;
1948         td->o.open_files = 0;
1949         td->files_index = 0;
1950 }
1951
1952 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1953 {
1954         int i;
1955
1956         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1957                 f->last_pos[i] = f->file_offset;
1958                 f->last_start[i] = -1ULL;
1959         }
1960
1961         if (fio_file_axmap(f))
1962                 axmap_reset(f->io_axmap);
1963         else if (fio_file_lfsr(f))
1964                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1965
1966         zbd_file_reset(td, f);
1967 }
1968
1969 bool fio_files_done(struct thread_data *td)
1970 {
1971         struct fio_file *f;
1972         unsigned int i;
1973
1974         for_each_file(td, f, i)
1975                 if (!fio_file_done(f))
1976                         return false;
1977
1978         return true;
1979 }
1980
1981 /* free memory used in initialization phase only */
1982 void filesetup_mem_free(void)
1983 {
1984         free_already_allocated();
1985 }
1986
1987 /*
1988  * This function is for platforms which support direct I/O but not O_DIRECT.
1989  */
1990 int fio_set_directio(struct thread_data *td, struct fio_file *f)
1991 {
1992 #ifdef FIO_OS_DIRECTIO
1993         int ret = fio_set_odirect(f);
1994
1995         if (ret) {
1996                 td_verror(td, ret, "fio_set_directio");
1997 #if defined(__sun__)
1998                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
1999                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2000                 } else {
2001                         log_err("fio: the file system does not seem to support direct IO\n");
2002                 }
2003 #else
2004                 log_err("fio: the file system does not seem to support direct IO\n");
2005 #endif
2006                 return -1;
2007         }
2008
2009         return 0;
2010 #else
2011         log_err("fio: direct IO is not supported on this host operating system\n");
2012         return -1;
2013 #endif
2014 }