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 || __e == EDQUOT) {
 230                                         const char *__e_name;
 231                                         if (td->o.fill_device)
 232                                                 break;
 233                                         if (__e == ENOSPC)
 234                                                 __e_name = "ENOSPC";
 235                                         else
 236                                                 __e_name = "EDQUOT";
 237                                         log_info("fio: %s on laying out "
 238                                                  "file, stopping\n", __e_name);
 239                                 }
 240                                 td_verror(td, errno, "write");
 241                         } else
 242                                 td_verror(td, EIO, "write");
 243
 244                         goto err;
 245                 }
 246         }
 247
 248         if (td->terminate) {
 249                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 250                 td_io_unlink_file(td, f);
 251         } else if (td->o.create_fsync) {
 252                 if (fsync(f->fd) < 0) {
 253                         td_verror(td, errno, "fsync");
 254                         goto err;
 255                 }
 256         }
 257         if (td->o.fill_device && !td_write(td)) {
 258                 fio_file_clear_size_known(f);
 259                 if (td_io_get_file_size(td, f))
 260                         goto err;
 261                 if (f->io_size > f->real_file_size)
 262                         f->io_size = f->real_file_size;
 263         }
 264
 265         free(b);
 266 done:
 267         return 0;
 268 err:
 269         close(f->fd);
 270         f->fd = -1;
 271         if (b)
 272                 free(b);
 273         return 1;
 274 }
 275
 276 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 277 {
 278         int r, did_open = 0, old_runstate;
 279         unsigned long long left;
 280         unsigned long long bs;
 281         bool ret = true;
 282         char *b;
 283
 284         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 285             td_ioengine_flagged(td, FIO_NOIO))
 286                 return true;
 287
 288         if (f->filetype == FIO_TYPE_CHAR)
 289                 return true;
 290
 291         if (!fio_file_open(f)) {
 292                 if (td->io_ops->open_file(td, f)) {
 293                         log_err("fio: cannot pre-read, failed to open file\n");
 294                         return false;
 295                 }
 296                 did_open = 1;
 297         }
 298
 299         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 300
 301         left = f->io_size;
 302         bs = td->o.max_bs[DDIR_READ];
 303         if (bs > left)
 304                 bs = left;
 305
 306         b = calloc(1, bs);
 307         if (!b) {
 308                 td_verror(td, errno, "malloc");
 309                 ret = false;
 310                 goto error;
 311         }
 312
 313         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 314                 td_verror(td, errno, "lseek");
 315                 log_err("fio: failed to lseek pre-read file\n");
 316                 ret = false;
 317                 goto error;
 318         }
 319
 320         while (left && !td->terminate) {
 321                 if (bs > left)
 322                         bs = left;
 323
 324                 r = read(f->fd, b, bs);
 325
 326                 if (r == (int) bs) {
 327                         left -= bs;
 328                         continue;
 329                 } else {
 330                         td_verror(td, EIO, "pre_read");
 331                         break;
 332                 }
 333         }
 334
 335 error:
 336         td_restore_runstate(td, old_runstate);
 337
 338         if (did_open)
 339                 td->io_ops->close_file(td, f);
 340
 341         free(b);
 342         return ret;
 343 }
 344
 345 /*
 346  * Generic function to prepopulate regular file with data.
 347  * Useful if you want to make sure I/O engine has data to read.
 348  * Leaves f->fd open on success, caller must close.
 349  */
 350 int generic_prepopulate_file(struct thread_data *td, struct fio_file *f)
 351 {
 352         int flags;
 353         unsigned long long left, bs;
 354         char *b = NULL;
 355
 356         /* generic function for regular files only */
 357         assert(f->filetype == FIO_TYPE_FILE);
 358
 359         if (read_only) {
 360                 log_err("fio: refusing to write a file due to read-only\n");
 361                 return 0;
 362         }
 363
 364         flags = O_WRONLY;
 365         if (td->o.allow_create)
 366                 flags |= O_CREAT;
 367
 368 #ifdef WIN32
 369         flags |= _O_BINARY;
 370 #endif
 371
 372         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 373         f->fd = open(f->file_name, flags, 0644);
 374         if (f->fd < 0) {
 375                 int err = errno;
 376
 377                 if (err == ENOENT && !td->o.allow_create)
 378                         log_err("fio: file creation disallowed by "
 379                                         "allow_file_create=0\n");
 380                 else
 381                         td_verror(td, err, "open");
 382                 return 1;
 383         }
 384
 385         left = f->real_file_size;
 386         bs = td->o.max_bs[DDIR_WRITE];
 387         if (bs > left)
 388                 bs = left;
 389
 390         b = malloc(bs);
 391         if (!b) {
 392                 td_verror(td, errno, "malloc");
 393                 goto err;
 394         }
 395
 396         while (left && !td->terminate) {
 397                 ssize_t r;
 398
 399                 if (bs > left)
 400                         bs = left;
 401
 402                 fill_io_buffer(td, b, bs, bs);
 403
 404                 r = write(f->fd, b, bs);
 405
 406                 if (r > 0) {
 407                         left -= r;
 408                 } else {
 409                         td_verror(td, errno, "write");
 410                         goto err;
 411                 }
 412         }
 413
 414         if (td->terminate) {
 415                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 416                 td_io_unlink_file(td, f);
 417         } else if (td->o.create_fsync) {
 418                 if (fsync(f->fd) < 0) {
 419                         td_verror(td, errno, "fsync");
 420                         goto err;
 421                 }
 422         }
 423
 424         free(b);
 425         return 0;
 426 err:
 427         close(f->fd);
 428         f->fd = -1;
 429         if (b)
 430                 free(b);
 431         return 1;
 432 }
 433
 434 unsigned long long get_rand_file_size(struct thread_data *td)
 435 {
 436         unsigned long long ret, sized;
 437         uint64_t frand_max;
 438         uint64_t r;
 439
 440         frand_max = rand_max(&td->file_size_state);
 441         r = __rand(&td->file_size_state);
 442         sized = td->o.file_size_high - td->o.file_size_low;
 443         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 444         ret += td->o.file_size_low;
 445         ret -= (ret % td->o.rw_min_bs);
 446         return ret;
 447 }
 448
 449 static int file_size(struct thread_data *td, struct fio_file *f)
 450 {
 451         struct stat st;
 452
 453         if (stat(f->file_name, &st) == -1) {
 454                 td_verror(td, errno, "fstat");
 455                 return 1;
 456         }
 457
 458         f->real_file_size = st.st_size;
 459         return 0;
 460 }
 461
 462 static int bdev_size(struct thread_data *td, struct fio_file *f)
 463 {
 464         unsigned long long bytes = 0;
 465         int r;
 466
 467         if (td->io_ops->open_file(td, f)) {
 468                 log_err("fio: failed opening blockdev %s for size check\n",
 469                         f->file_name);
 470                 return 1;
 471         }
 472
 473         r = blockdev_size(f, &bytes);
 474         if (r) {
 475                 td_verror(td, r, "blockdev_size");
 476                 goto err;
 477         }
 478
 479         if (!bytes) {
 480                 log_err("%s: zero sized block device?\n", f->file_name);
 481                 goto err;
 482         }
 483
 484         f->real_file_size = bytes;
 485         td->io_ops->close_file(td, f);
 486         return 0;
 487 err:
 488         td->io_ops->close_file(td, f);
 489         return 1;
 490 }
 491
 492 static int char_size(struct thread_data *td, struct fio_file *f)
 493 {
 494 #ifdef FIO_HAVE_CHARDEV_SIZE
 495         unsigned long long bytes = 0;
 496         int r;
 497
 498         if (td->io_ops->open_file(td, f)) {
 499                 log_err("fio: failed opening chardev %s for size check\n",
 500                         f->file_name);
 501                 return 1;
 502         }
 503
 504         r = chardev_size(f, &bytes);
 505         if (r) {
 506                 td_verror(td, r, "chardev_size");
 507                 goto err;
 508         }
 509
 510         if (!bytes) {
 511                 log_err("%s: zero sized char device?\n", f->file_name);
 512                 goto err;
 513         }
 514
 515         f->real_file_size = bytes;
 516         td->io_ops->close_file(td, f);
 517         return 0;
 518 err:
 519         td->io_ops->close_file(td, f);
 520         return 1;
 521 #else
 522         f->real_file_size = -1ULL;
 523         return 0;
 524 #endif
 525 }
 526
 527 static int get_file_size(struct thread_data *td, struct fio_file *f)
 528 {
 529         int ret = 0;
 530
 531         if (fio_file_size_known(f))
 532                 return 0;
 533
 534         if (f->filetype == FIO_TYPE_FILE)
 535                 ret = file_size(td, f);
 536         else if (f->filetype == FIO_TYPE_BLOCK)
 537                 ret = bdev_size(td, f);
 538         else if (f->filetype == FIO_TYPE_CHAR)
 539                 ret = char_size(td, f);
 540         else {
 541                 f->real_file_size = -1;
 542                 log_info("%s: failed to get file size of %s\n", td->o.name,
 543                                         f->file_name);
 544                 return 1; /* avoid offset extends end error message */
 545         }
 546
 547         /*
 548          * Leave ->real_file_size with 0 since it could be expectation
 549          * of initial setup for regular files.
 550          */
 551         if (ret)
 552                 return ret;
 553
 554         /*
 555          * ->file_offset normally hasn't been initialized yet, so this
 556          * is basically always false unless ->real_file_size is -1, but
 557          * if ->real_file_size is -1 this message doesn't make sense.
 558          * As a result, this message is basically useless.
 559          */
 560         if (f->file_offset > f->real_file_size) {
 561                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 562                                         (unsigned long long) f->file_offset,
 563                                         (unsigned long long) f->real_file_size);
 564                 return 1;
 565         }
 566
 567         fio_file_set_size_known(f);
 568         return 0;
 569 }
 570
 571 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 572                                    unsigned long long off,
 573                                    unsigned long long len)
 574 {
 575         int errval = 0, ret = 0;
 576
 577 #ifdef CONFIG_ESX
 578         return 0;
 579 #endif
 580
 581         if (len == -1ULL)
 582                 len = f->io_size;
 583         if (off == -1ULL)
 584                 off = f->file_offset;
 585
 586         if (len == -1ULL || off == -1ULL)
 587                 return 0;
 588
 589         if (td->io_ops->invalidate) {
 590                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 591                         f->file_name);
 592                 ret = td->io_ops->invalidate(td, f);
 593                 if (ret < 0)
 594                         errval = -ret;
 595         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 596                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
 597                        td->io_ops->name);
 598         } else if (f->filetype == FIO_TYPE_FILE) {
 599                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 600                         f->file_name, off, len);
 601                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 602                 if (ret)
 603                         errval = ret;
 604         } else if (f->filetype == FIO_TYPE_BLOCK) {
 605                 int retry_count = 0;
 606
 607                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 608                 ret = blockdev_invalidate_cache(f);
 609                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 610                         /*
 611                          * Linux multipath devices reject ioctl while
 612                          * the maps are being updated. That window can
 613                          * last tens of milliseconds; we'll try up to
 614                          * a quarter of a second.
 615                          */
 616                         usleep(10000);
 617                         ret = blockdev_invalidate_cache(f);
 618                 }
 619                 if (ret < 0 && errno == EACCES && geteuid()) {
 620                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 621                                 log_err("fio: only root may flush block "
 622                                         "devices. Cache flush bypassed!\n");
 623                         }
 624                 }
 625                 if (ret < 0)
 626                         errval = errno;
 627         } else if (f->filetype == FIO_TYPE_CHAR ||
 628                    f->filetype == FIO_TYPE_PIPE) {
 629                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 630         }
 631
 632         /*
 633          * Cache flushing isn't a fatal condition, and we know it will
 634          * happen on some platforms where we don't have the proper
 635          * function to flush eg block device caches. So just warn and
 636          * continue on our way.
 637          */
 638         if (errval)
 639                 log_info("fio: cache invalidation of %s failed: %s\n",
 640                          f->file_name, strerror(errval));
 641
 642         return 0;
 643
 644 }
 645
 646 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 647 {
 648         if (!fio_file_open(f))
 649                 return 0;
 650
 651         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 652 }
 653
 654 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 655 {
 656         int ret = 0;
 657
 658         dprint(FD_FILE, "fd close %s\n", f->file_name);
 659
 660         remove_file_hash(f);
 661
 662         if (close(f->fd) < 0)
 663                 ret = errno;
 664
 665         f->fd = -1;
 666
 667         if (f->shadow_fd != -1) {
 668                 close(f->shadow_fd);
 669                 f->shadow_fd = -1;
 670         }
 671
 672         f->engine_pos = 0;
 673         return ret;
 674 }
 675
 676 int file_lookup_open(struct fio_file *f, int flags)
 677 {
 678         struct fio_file *__f;
 679         int from_hash;
 680
 681         __f = lookup_file_hash(f->file_name);
 682         if (__f) {
 683                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 684                 f->lock = __f->lock;
 685                 from_hash = 1;
 686         } else {
 687                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 688                 from_hash = 0;
 689         }
 690
 691 #ifdef WIN32
 692         flags |= _O_BINARY;
 693 #endif
 694
 695         f->fd = open(f->file_name, flags, 0600);
 696         return from_hash;
 697 }
 698
 699 static int file_close_shadow_fds(struct thread_data *td)
 700 {
 701         struct fio_file *f;
 702         int num_closed = 0;
 703         unsigned int i;
 704
 705         for_each_file(td, f, i) {
 706                 if (f->shadow_fd == -1)
 707                         continue;
 708
 709                 close(f->shadow_fd);
 710                 f->shadow_fd = -1;
 711                 num_closed++;
 712         }
 713
 714         return num_closed;
 715 }
 716
 717 int generic_open_file(struct thread_data *td, struct fio_file *f)
 718 {
 719         int is_std = 0;
 720         int flags = 0;
 721         int from_hash = 0;
 722
 723         dprint(FD_FILE, "fd open %s\n", f->file_name);
 724
 725         if (!strcmp(f->file_name, "-")) {
 726                 if (td_rw(td)) {
 727                         log_err("fio: can't read/write to stdin/out\n");
 728                         return 1;
 729                 }
 730                 is_std = 1;
 731
 732                 /*
 733                  * move output logging to stderr, if we are writing to stdout
 734                  */
 735                 if (td_write(td))
 736                         f_out = stderr;
 737         }
 738
 739         if (td->o.odirect)
 740                 flags |= OS_O_DIRECT;
 741         flags |= td->o.sync_io;
 742         if (td->o.create_on_open && td->o.allow_create)
 743                 flags |= O_CREAT;
 744         if (f->filetype != FIO_TYPE_FILE)
 745                 flags |= FIO_O_NOATIME;
 746
 747 open_again:
 748         if (td_write(td)) {
 749                 if (!read_only)
 750                         flags |= O_RDWR;
 751
 752                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 753                         flags |= O_CREAT;
 754
 755                 if (is_std)
 756                         f->fd = dup(STDOUT_FILENO);
 757                 else
 758                         from_hash = file_lookup_open(f, flags);
 759         } else if (td_read(td)) {
 760                 if (td_ioengine_flagged(td, FIO_RO_NEEDS_RW_OPEN) && !read_only)
 761                         flags |= O_RDWR;
 762                 else
 763                         flags |= O_RDONLY;
 764
 765                 if (is_std)
 766                         f->fd = dup(STDIN_FILENO);
 767                 else
 768                         from_hash = file_lookup_open(f, flags);
 769         } else if (td_trim(td)) {
 770                 assert(!td_rw(td)); /* should have matched above */
 771                 if (!read_only)
 772                         flags |= O_RDWR;
 773                 from_hash = file_lookup_open(f, flags);
 774         }
 775
 776         if (f->fd == -1) {
 777                 char buf[FIO_VERROR_SIZE];
 778                 int __e = errno;
 779
 780                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 781                         flags &= ~FIO_O_NOATIME;
 782                         goto open_again;
 783                 }
 784                 if (__e == EMFILE && file_close_shadow_fds(td))
 785                         goto open_again;
 786
 787                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 788
 789                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 790                         log_err("fio: looks like your file system does not " \
 791                                 "support direct=1/buffered=0\n");
 792                 }
 793
 794                 td_verror(td, __e, buf);
 795                 return 1;
 796         }
 797
 798         if (!from_hash && f->fd != -1) {
 799                 if (add_file_hash(f)) {
 800                         int fio_unused ret;
 801
 802                         /*
 803                          * Stash away descriptor for later close. This is to
 804                          * work-around a "feature" on Linux, where a close of
 805                          * an fd that has been opened for write will trigger
 806                          * udev to call blkid to check partitions, fs id, etc.
 807                          * That pollutes the device cache, which can slow down
 808                          * unbuffered accesses.
 809                          */
 810                         if (f->shadow_fd == -1)
 811                                 f->shadow_fd = f->fd;
 812                         else {
 813                                 /*
 814                                  * OK to ignore, we haven't done anything
 815                                  * with it
 816                                  */
 817                                 ret = generic_close_file(td, f);
 818                         }
 819                         goto open_again;
 820                 }
 821         }
 822
 823         return 0;
 824 }
 825
 826 /*
 827  * This function i.e. get_file_size() is the default .get_file_size
 828  * implementation of majority of I/O engines.
 829  */
 830 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 831 {
 832         return get_file_size(td, f);
 833 }
 834
 835 /*
 836  * open/close all files, so that ->real_file_size gets set
 837  */
 838 static int get_file_sizes(struct thread_data *td)
 839 {
 840         struct fio_file *f;
 841         unsigned int i;
 842         int err = 0;
 843
 844         for_each_file(td, f, i) {
 845                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 846                                                                 f->file_name);
 847
 848                 if (td_io_get_file_size(td, f)) {
 849                         if (td->error != ENOENT) {
 850                                 log_err("%s\n", td->verror);
 851                                 err = 1;
 852                                 break;
 853                         }
 854                         clear_error(td);
 855                 }
 856
 857                 /*
 858                  * There are corner cases where we end up with -1 for
 859                  * ->real_file_size due to unsupported file type, etc.
 860                  * We then just set to size option value divided by number
 861                  * of files, similar to the way file ->io_size is set.
 862                  * stat(2) failure doesn't set ->real_file_size to -1.
 863                  */
 864                 if (f->real_file_size == -1ULL && td->o.size)
 865                         f->real_file_size = td->o.size / td->o.nr_files;
 866         }
 867
 868         return err;
 869 }
 870
 871 struct fio_mount {
 872         struct flist_head list;
 873         const char *base;
 874         char __base[256];
 875         unsigned int key;
 876 };
 877
 878 /*
 879  * Get free number of bytes for each file on each unique mount.
 880  */
 881 static unsigned long long get_fs_free_counts(struct thread_data *td)
 882 {
 883         struct flist_head *n, *tmp;
 884         unsigned long long ret = 0;
 885         struct fio_mount *fm;
 886         FLIST_HEAD(list);
 887         struct fio_file *f;
 888         unsigned int i;
 889
 890         for_each_file(td, f, i) {
 891                 struct stat sb;
 892                 char buf[256];
 893
 894                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 895                         if (f->real_file_size != -1ULL)
 896                                 ret += f->real_file_size;
 897                         continue;
 898                 } else if (f->filetype != FIO_TYPE_FILE)
 899                         continue;
 900
 901                 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
 902
 903                 if (stat(buf, &sb) < 0) {
 904                         if (errno != ENOENT)
 905                                 break;
 906                         strcpy(buf, ".");
 907                         if (stat(buf, &sb) < 0)
 908                                 break;
 909                 }
 910
 911                 fm = NULL;
 912                 flist_for_each(n, &list) {
 913                         fm = flist_entry(n, struct fio_mount, list);
 914                         if (fm->key == sb.st_dev)
 915                                 break;
 916
 917                         fm = NULL;
 918                 }
 919
 920                 if (fm)
 921                         continue;
 922
 923                 fm = calloc(1, sizeof(*fm));
 924                 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
 925                 fm->base = basename(fm->__base);
 926                 fm->key = sb.st_dev;
 927                 flist_add(&fm->list, &list);
 928         }
 929
 930         flist_for_each_safe(n, tmp, &list) {
 931                 unsigned long long sz;
 932
 933                 fm = flist_entry(n, struct fio_mount, list);
 934                 flist_del(&fm->list);
 935
 936                 sz = get_fs_free_size(fm->base);
 937                 if (sz && sz != -1ULL)
 938                         ret += sz;
 939
 940                 free(fm);
 941         }
 942
 943         return ret;
 944 }
 945
 946 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 947 {
 948         bool align = false;
 949         struct thread_options *o = &td->o;
 950         unsigned long long align_bs;
 951         unsigned long long offset;
 952         unsigned long long increment;
 953
 954         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 955                 return f->real_file_size;
 956
 957         if (o->offset_increment_percent) {
 958                 assert(!o->offset_increment);
 959                 increment = o->offset_increment_percent * f->real_file_size / 100;
 960                 align = true;
 961         } else
 962                 increment = o->offset_increment;
 963
 964         if (o->start_offset_percent > 0) {
 965                 /* calculate the raw offset */
 966                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 967                         (td->subjob_number * increment);
 968
 969                 align = true;
 970         } else {
 971                 /* start_offset_percent not set */
 972                 offset = o->start_offset +
 973                                 td->subjob_number * increment;
 974         }
 975
 976         if (align) {
 977                 /*
 978                  * if offset_align is provided, use it
 979                  */
 980                 if (fio_option_is_set(o, start_offset_align)) {
 981                         align_bs = o->start_offset_align;
 982                 } else {
 983                         /* else take the minimum block size */
 984                         align_bs = td_min_bs(td);
 985                 }
 986
 987                 /*
 988                  * block align the offset at the next available boundary at
 989                  * ceiling(offset / align_bs) * align_bs
 990                  */
 991                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 992         }
 993
 994         return offset;
 995 }
 996
 997 /*
 998  * Find longest path component that exists and return its length
 999  */
1000 int longest_existing_path(char *path) {
1001         char buf[PATH_MAX];
1002         bool done;
1003         char *buf_pos;
1004         int offset;
1005 #ifdef WIN32
1006         DWORD dwAttr;
1007 #else
1008         struct stat sb;
1009 #endif
1010
1011         sprintf(buf, "%s", path);
1012         done = false;
1013         while (!done) {
1014                 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1015                 if (!buf_pos) {
1016                         offset = 0;
1017                         break;
1018                 }
1019
1020                 *(buf_pos + 1) = '\0';
1021
1022 #ifdef WIN32
1023                 dwAttr = GetFileAttributesA(buf);
1024                 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1025                         done = true;
1026                 }
1027 #else
1028                 if (stat(buf, &sb) == 0)
1029                         done = true;
1030 #endif
1031                 if (done)
1032                         offset = buf_pos - buf;
1033                 else
1034                         *buf_pos = '\0';
1035         }
1036
1037         return offset;
1038 }
1039
1040 static bool create_work_dirs(struct thread_data *td, const char *fname)
1041 {
1042         char path[PATH_MAX];
1043         char *start, *end;
1044         int offset;
1045
1046         snprintf(path, PATH_MAX, "%s", fname);
1047         start = path;
1048
1049         offset = longest_existing_path(path);
1050         end = start + offset;
1051         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1052                 if (end == start) {
1053                         end++;
1054                         continue;
1055                 }
1056                 *end = '\0';
1057                 errno = 0;
1058                 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1059                         log_err("fio: failed to create dir (%s): %s\n",
1060                                 start, strerror(errno));
1061                         return false;
1062                 }
1063                 *end = FIO_OS_PATH_SEPARATOR;
1064                 end++;
1065         }
1066         td->flags |= TD_F_DIRS_CREATED;
1067         return true;
1068 }
1069
1070 /*
1071  * Open the files and setup files sizes, creating files if necessary.
1072  */
1073 int setup_files(struct thread_data *td)
1074 {
1075         unsigned long long total_size, extend_size;
1076         struct thread_options *o = &td->o;
1077         struct fio_file *f;
1078         unsigned int i, nr_fs_extra = 0;
1079         int err = 0, need_extend;
1080         int old_state;
1081         const unsigned long long bs = td_min_bs(td);
1082         uint64_t fs = 0;
1083
1084         dprint(FD_FILE, "setup files\n");
1085
1086         old_state = td_bump_runstate(td, TD_SETTING_UP);
1087
1088         for_each_file(td, f, i) {
1089                 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1090                     strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1091                     !(td->flags & TD_F_DIRS_CREATED) &&
1092                     !create_work_dirs(td, f->file_name))
1093                         goto err_out;
1094         }
1095
1096         /*
1097          * Find out physical size of files or devices for this thread,
1098          * before we determine I/O size and range of our targets.
1099          * If ioengine defines a setup() method, it's responsible for
1100          * opening the files and setting f->real_file_size to indicate
1101          * the valid range for that file.
1102          */
1103         if (td->io_ops->setup)
1104                 err = td->io_ops->setup(td);
1105         else
1106                 err = get_file_sizes(td);
1107
1108         if (err)
1109                 goto err_out;
1110
1111         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1112                 err = zbd_init_files(td);
1113                 if (err)
1114                         goto err_out;
1115         }
1116         zbd_recalc_options_with_zone_granularity(td);
1117
1118         if (o->read_iolog_file)
1119                 goto done;
1120
1121         /*
1122          * check sizes. if the files/devices do not exist and the size
1123          * isn't passed to fio, abort.
1124          */
1125         total_size = 0;
1126         for_each_file(td, f, i) {
1127                 f->fileno = i;
1128                 if (f->real_file_size == -1ULL)
1129                         total_size = -1ULL;
1130                 else
1131                         total_size += f->real_file_size;
1132         }
1133
1134         if (o->fill_device)
1135                 td->fill_device_size = get_fs_free_counts(td);
1136
1137         /*
1138          * device/file sizes are zero and no size given, punt
1139          */
1140         if ((!total_size || total_size == -1ULL) && !o->size &&
1141             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1142             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1143                 log_err("%s: you need to specify size=\n", o->name);
1144                 td_verror(td, EINVAL, "total_file_size");
1145                 goto err_out;
1146         }
1147
1148         /*
1149          * Calculate per-file size and potential extra size for the
1150          * first files, if needed (i.e. if we don't have a fixed size).
1151          */
1152         if (!o->file_size_low && o->nr_files) {
1153                 uint64_t all_fs;
1154
1155                 fs = o->size / o->nr_files;
1156                 all_fs = fs * o->nr_files;
1157
1158                 if (all_fs < o->size)
1159                         nr_fs_extra = (o->size - all_fs) / bs;
1160         }
1161
1162         /*
1163          * now file sizes are known, so we can set ->io_size. if size= is
1164          * not given, ->io_size is just equal to ->real_file_size. if size
1165          * is given, ->io_size is size / nr_files.
1166          */
1167         extend_size = total_size = 0;
1168         need_extend = 0;
1169         for_each_file(td, f, i) {
1170                 f->file_offset = get_start_offset(td, f);
1171
1172                 /*
1173                  * Update ->io_size depending on options specified.
1174                  * ->file_size_low being 0 means filesize option isn't set.
1175                  * Non zero ->file_size_low equals ->file_size_high means
1176                  * filesize option is set in a fixed size format.
1177                  * Non zero ->file_size_low not equals ->file_size_high means
1178                  * filesize option is set in a range format.
1179                  */
1180                 if (!o->file_size_low) {
1181                         /*
1182                          * no file size or range given, file size is equal to
1183                          * total size divided by number of files. If the size
1184                          * doesn't divide nicely with the min blocksize,
1185                          * make the first files bigger.
1186                          */
1187                         f->io_size = fs;
1188                         if (nr_fs_extra) {
1189                                 nr_fs_extra--;
1190                                 f->io_size += bs;
1191                         }
1192
1193                         /*
1194                          * We normally don't come here for regular files, but
1195                          * if the result is 0 for a regular file, set it to the
1196                          * real file size. This could be size of the existing
1197                          * one if it already exists, but otherwise will be set
1198                          * to 0. A new file won't be created because
1199                          * ->io_size + ->file_offset equals ->real_file_size.
1200                          */
1201                         if (!f->io_size) {
1202                                 if (f->file_offset > f->real_file_size)
1203                                         goto err_offset;
1204                                 f->io_size = f->real_file_size - f->file_offset;
1205                                 if (!f->io_size)
1206                                         log_info("fio: file %s may be ignored\n",
1207                                                 f->file_name);
1208                         }
1209                 } else if (f->real_file_size < o->file_size_low ||
1210                            f->real_file_size > o->file_size_high) {
1211                         if (f->file_offset > o->file_size_low)
1212                                 goto err_offset;
1213                         /*
1214                          * file size given. if it's fixed, use that. if it's a
1215                          * range, generate a random size in-between.
1216                          */
1217                         if (o->file_size_low == o->file_size_high)
1218                                 f->io_size = o->file_size_low - f->file_offset;
1219                         else {
1220                                 f->io_size = get_rand_file_size(td)
1221                                                 - f->file_offset;
1222                         }
1223                 } else
1224                         f->io_size = f->real_file_size - f->file_offset;
1225
1226                 if (f->io_size == -1ULL)
1227                         total_size = -1ULL;
1228                 else {
1229                         uint64_t io_size;
1230
1231                         if (o->size_percent && o->size_percent != 100) {
1232                                 uint64_t file_size;
1233
1234                                 file_size = f->io_size + f->file_offset;
1235                                 f->io_size = (file_size *
1236                                               o->size_percent) / 100;
1237                                 if (f->io_size > (file_size - f->file_offset))
1238                                         f->io_size = file_size - f->file_offset;
1239
1240                                 f->io_size -= (f->io_size % td_min_bs(td));
1241                         }
1242
1243                         io_size = f->io_size;
1244                         if (o->io_size_percent && o->io_size_percent != 100) {
1245                                 io_size *= o->io_size_percent;
1246                                 io_size /= 100;
1247                         }
1248
1249                         total_size += io_size;
1250                 }
1251
1252                 if (f->filetype == FIO_TYPE_FILE &&
1253                     (f->io_size + f->file_offset) > f->real_file_size) {
1254                         if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1255                             !o->create_on_open) {
1256                                 need_extend++;
1257                                 extend_size += (f->io_size + f->file_offset);
1258                                 fio_file_set_extend(f);
1259                         } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1260                                    (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1261                                     td_ioengine_flagged(td, FIO_FAKEIO)))
1262                                 f->real_file_size = f->io_size + f->file_offset;
1263                 }
1264         }
1265
1266         if (td->o.block_error_hist) {
1267                 int len;
1268
1269                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1270                 f = td->files[0];
1271                 len = f->io_size / td->o.bs[DDIR_TRIM];
1272                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1273                         log_err("fio: cannot calculate block histogram with "
1274                                 "%d trim blocks, maximum %d\n",
1275                                 len, MAX_NR_BLOCK_INFOS);
1276                         td_verror(td, EINVAL, "block_error_hist");
1277                         goto err_out;
1278                 }
1279
1280                 td->ts.nr_block_infos = len;
1281                 for (i = 0; i < len; i++)
1282                         td->ts.block_infos[i] =
1283                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1284         } else
1285                 td->ts.nr_block_infos = 0;
1286
1287         if (!o->size || (total_size && o->size > total_size))
1288                 o->size = total_size;
1289
1290         if (o->size < td_min_bs(td)) {
1291                 log_err("fio: blocksize is larger than data set range\n");
1292                 goto err_out;
1293         }
1294
1295         /*
1296          * See if we need to extend some files, typically needed when our
1297          * target regular files don't exist yet, but our jobs require them
1298          * initially due to read I/Os.
1299          */
1300         if (need_extend) {
1301                 temp_stall_ts = 1;
1302                 if (output_format & FIO_OUTPUT_NORMAL) {
1303                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1304                                  o->name,
1305                                  need_extend > 1 ? "s" : "",
1306                                  need_extend,
1307                                  need_extend > 1 ? "s" : "",
1308                                  need_extend > 1 ? "total " : "",
1309                                  extend_size >> 20);
1310                 }
1311
1312                 for_each_file(td, f, i) {
1313                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1314
1315                         if (!fio_file_extend(f))
1316                                 continue;
1317
1318                         assert(f->filetype == FIO_TYPE_FILE);
1319                         fio_file_clear_extend(f);
1320                         if (!o->fill_device) {
1321                                 old_len = f->real_file_size;
1322                                 extend_len = f->io_size + f->file_offset -
1323                                                 old_len;
1324                         }
1325                         f->real_file_size = (f->io_size + f->file_offset);
1326                         err = extend_file(td, f);
1327                         if (err)
1328                                 break;
1329
1330                         err = __file_invalidate_cache(td, f, old_len,
1331                                                                 extend_len);
1332
1333                         /*
1334                          * Shut up static checker
1335                          */
1336                         if (f->fd != -1)
1337                                 close(f->fd);
1338
1339                         f->fd = -1;
1340                         if (err)
1341                                 break;
1342                 }
1343                 temp_stall_ts = 0;
1344         }
1345
1346         if (err)
1347                 goto err_out;
1348
1349         /*
1350          * Prepopulate files with data. It might be expected to read some
1351          * "real" data instead of zero'ed files (if no writes to file occurred
1352          * prior to a read job). Engine has to provide a way to do that.
1353          */
1354         if (td->io_ops->prepopulate_file) {
1355                 temp_stall_ts = 1;
1356
1357                 for_each_file(td, f, i) {
1358                         if (output_format & FIO_OUTPUT_NORMAL) {
1359                                 log_info("%s: Prepopulating IO file (%s)\n",
1360                                                         o->name, f->file_name);
1361                         }
1362
1363                         err = td->io_ops->prepopulate_file(td, f);
1364                         if (err)
1365                                 break;
1366
1367                         err = __file_invalidate_cache(td, f, f->file_offset,
1368                                                                 f->io_size);
1369
1370                         /*
1371                          * Shut up static checker
1372                          */
1373                         if (f->fd != -1)
1374                                 close(f->fd);
1375
1376                         f->fd = -1;
1377                         if (err)
1378                                 break;
1379                 }
1380                 temp_stall_ts = 0;
1381         }
1382
1383         if (err)
1384                 goto err_out;
1385
1386         /*
1387          * iolog already set the total io size, if we read back
1388          * stored entries.
1389          */
1390         if (!o->read_iolog_file) {
1391                 if (o->io_size)
1392                         td->total_io_size = o->io_size * o->loops;
1393                 else
1394                         td->total_io_size = o->size * o->loops;
1395         }
1396
1397 done:
1398         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1399                 err = zbd_setup_files(td);
1400                 if (err)
1401                         goto err_out;
1402         }
1403
1404         if (o->create_only)
1405                 td->done = 1;
1406
1407         td_restore_runstate(td, old_state);
1408
1409         if (td->o.fdp) {
1410                 err = fdp_init(td);
1411                 if (err)
1412                         goto err_out;
1413         }
1414
1415         return 0;
1416
1417 err_offset:
1418         log_err("%s: you need to specify valid offset=\n", o->name);
1419 err_out:
1420         td_restore_runstate(td, old_state);
1421         return 1;
1422 }
1423
1424 bool pre_read_files(struct thread_data *td)
1425 {
1426         struct fio_file *f;
1427         unsigned int i;
1428
1429         dprint(FD_FILE, "pre_read files\n");
1430
1431         for_each_file(td, f, i) {
1432                 if (!pre_read_file(td, f))
1433                         return false;
1434         }
1435
1436         return true;
1437 }
1438
1439 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1440 {
1441         unsigned int range_size, seed;
1442         uint64_t nranges;
1443         uint64_t fsize;
1444
1445         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1446         fsize = min(f->real_file_size, f->io_size);
1447
1448         nranges = (fsize + range_size - 1ULL) / range_size;
1449
1450         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number *
1451                 td->rand_seeds[FIO_RAND_BLOCK_OFF];
1452
1453         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1454                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1455         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1456                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1457         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1458                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1459 }
1460
1461 static bool init_rand_distribution(struct thread_data *td)
1462 {
1463         struct fio_file *f;
1464         unsigned int i;
1465         int state;
1466
1467         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1468             td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1469             td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1470                 return false;
1471
1472         state = td_bump_runstate(td, TD_SETTING_UP);
1473
1474         for_each_file(td, f, i)
1475                 __init_rand_distribution(td, f);
1476
1477         td_restore_runstate(td, state);
1478         return true;
1479 }
1480
1481 /*
1482  * Check if the number of blocks exceeds the randomness capability of
1483  * the selected generator. Tausworthe is 32-bit, the others are fully
1484  * 64-bit capable.
1485  */
1486 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1487                                  uint64_t blocks)
1488 {
1489         if (blocks <= FRAND32_MAX)
1490                 return 0;
1491         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1492                 return 0;
1493
1494         /*
1495          * If the user hasn't specified a random generator, switch
1496          * to tausworthe64 with informational warning. If the user did
1497          * specify one, just warn.
1498          */
1499         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1500                         f->file_name);
1501
1502         if (!fio_option_is_set(&td->o, random_generator)) {
1503                 log_info("fio: Switching to tausworthe64. Use the "
1504                          "random_generator= option to get rid of this "
1505                          "warning.\n");
1506                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1507                 return 0;
1508         }
1509
1510         /*
1511          * Just make this information to avoid breaking scripts.
1512          */
1513         log_info("fio: Use the random_generator= option to switch to lfsr or "
1514                          "tausworthe64.\n");
1515         return 0;
1516 }
1517
1518 bool init_random_map(struct thread_data *td)
1519 {
1520         unsigned long long blocks;
1521         struct fio_file *f;
1522         unsigned int i;
1523
1524         if (init_rand_distribution(td))
1525                 return true;
1526         if (!td_random(td))
1527                 return true;
1528
1529         for_each_file(td, f, i) {
1530                 uint64_t fsize = min(f->real_file_size, f->io_size);
1531
1532                 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1533                         fsize = td->o.zone_range;
1534
1535                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1536
1537                 if (check_rand_gen_limits(td, f, blocks))
1538                         return false;
1539
1540                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1541                         uint64_t seed;
1542
1543                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1544
1545                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1546                                 fio_file_set_lfsr(f);
1547                                 continue;
1548                         } else {
1549                                 log_err("fio: failed initializing LFSR\n");
1550                                 return false;
1551                         }
1552                 } else if (!td->o.norandommap) {
1553                         f->io_axmap = axmap_new(blocks);
1554                         if (f->io_axmap) {
1555                                 fio_file_set_axmap(f);
1556                                 continue;
1557                         }
1558                 } else if (td->o.norandommap)
1559                         continue;
1560
1561                 if (!td->o.softrandommap) {
1562                         log_err("fio: failed allocating random map. If running"
1563                                 " a large number of jobs, try the 'norandommap'"
1564                                 " option or set 'softrandommap'. Or give"
1565                                 " a larger --alloc-size to fio.\n");
1566                         return false;
1567                 }
1568
1569                 log_info("fio: file %s failed allocating random map. Running "
1570                          "job without.\n", f->file_name);
1571         }
1572
1573         return true;
1574 }
1575
1576 void close_files(struct thread_data *td)
1577 {
1578         struct fio_file *f;
1579         unsigned int i;
1580
1581         for_each_file(td, f, i) {
1582                 if (fio_file_open(f))
1583                         td_io_close_file(td, f);
1584         }
1585 }
1586
1587 void fio_file_free(struct fio_file *f)
1588 {
1589         if (fio_file_axmap(f))
1590                 axmap_free(f->io_axmap);
1591         if (f->ruhs_info)
1592                 sfree(f->ruhs_info);
1593         if (!fio_file_smalloc(f)) {
1594                 free(f->file_name);
1595                 free(f);
1596         } else {
1597                 sfree(f->file_name);
1598                 sfree(f);
1599         }
1600 }
1601
1602 void close_and_free_files(struct thread_data *td)
1603 {
1604         struct fio_file *f;
1605         unsigned int i;
1606
1607         dprint(FD_FILE, "close files\n");
1608
1609         for_each_file(td, f, i) {
1610                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1611                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1612                         td_io_unlink_file(td, f);
1613                 }
1614
1615                 if (fio_file_open(f))
1616                         td_io_close_file(td, f);
1617
1618                 remove_file_hash(f);
1619
1620                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1621                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1622                         td_io_unlink_file(td, f);
1623                 }
1624
1625                 zbd_close_file(f);
1626                 fdp_free_ruhs_info(f);
1627                 fio_file_free(f);
1628         }
1629
1630         td->o.filename = NULL;
1631         free(td->files);
1632         free(td->file_locks);
1633         td->files_index = 0;
1634         td->files = NULL;
1635         td->file_locks = NULL;
1636         td->o.file_lock_mode = FILE_LOCK_NONE;
1637         td->o.nr_files = 0;
1638 }
1639
1640 static void get_file_type(struct fio_file *f)
1641 {
1642         struct stat sb;
1643
1644         if (!strcmp(f->file_name, "-"))
1645                 f->filetype = FIO_TYPE_PIPE;
1646         else
1647                 f->filetype = FIO_TYPE_FILE;
1648
1649 #ifdef WIN32
1650         /* \\.\ is the device namespace in Windows, where every file is
1651          * a block device */
1652         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1653                 f->filetype = FIO_TYPE_BLOCK;
1654 #endif
1655
1656         if (!stat(f->file_name, &sb)) {
1657                 if (S_ISBLK(sb.st_mode))
1658                         f->filetype = FIO_TYPE_BLOCK;
1659                 else if (S_ISCHR(sb.st_mode))
1660                         f->filetype = FIO_TYPE_CHAR;
1661                 else if (S_ISFIFO(sb.st_mode))
1662                         f->filetype = FIO_TYPE_PIPE;
1663         }
1664 }
1665
1666 static bool __is_already_allocated(const char *fname, bool set)
1667 {
1668         struct flist_head *entry;
1669         bool ret;
1670
1671         ret = file_bloom_exists(fname, set);
1672         if (!ret)
1673                 return ret;
1674
1675         flist_for_each(entry, &filename_list) {
1676                 struct file_name *fn;
1677
1678                 fn = flist_entry(entry, struct file_name, list);
1679
1680                 if (!strcmp(fn->filename, fname))
1681                         return true;
1682         }
1683
1684         return false;
1685 }
1686
1687 static bool is_already_allocated(const char *fname)
1688 {
1689         bool ret;
1690
1691         fio_file_hash_lock();
1692         ret = __is_already_allocated(fname, false);
1693         fio_file_hash_unlock();
1694
1695         return ret;
1696 }
1697
1698 static void set_already_allocated(const char *fname)
1699 {
1700         struct file_name *fn;
1701
1702         fn = malloc(sizeof(struct file_name));
1703         fn->filename = strdup(fname);
1704
1705         fio_file_hash_lock();
1706         if (!__is_already_allocated(fname, true)) {
1707                 flist_add_tail(&fn->list, &filename_list);
1708                 fn = NULL;
1709         }
1710         fio_file_hash_unlock();
1711
1712         if (fn) {
1713                 free(fn->filename);
1714                 free(fn);
1715         }
1716 }
1717
1718 static void free_already_allocated(void)
1719 {
1720         struct flist_head *entry, *tmp;
1721         struct file_name *fn;
1722
1723         if (flist_empty(&filename_list))
1724                 return;
1725
1726         fio_file_hash_lock();
1727         flist_for_each_safe(entry, tmp, &filename_list) {
1728                 fn = flist_entry(entry, struct file_name, list);
1729                 free(fn->filename);
1730                 flist_del(&fn->list);
1731                 free(fn);
1732         }
1733
1734         fio_file_hash_unlock();
1735 }
1736
1737 static struct fio_file *alloc_new_file(struct thread_data *td)
1738 {
1739         struct fio_file *f;
1740
1741         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1742                 f = calloc(1, sizeof(*f));
1743         else
1744                 f = scalloc(1, sizeof(*f));
1745         if (!f) {
1746                 assert(0);
1747                 return NULL;
1748         }
1749
1750         f->fd = -1;
1751         f->shadow_fd = -1;
1752         fio_file_reset(td, f);
1753         if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1754                 fio_file_set_smalloc(f);
1755         return f;
1756 }
1757
1758 bool exists_and_not_regfile(const char *filename)
1759 {
1760         struct stat sb;
1761
1762         if (lstat(filename, &sb) == -1)
1763                 return false;
1764
1765 #ifndef WIN32 /* NOT Windows */
1766         if (S_ISREG(sb.st_mode))
1767                 return false;
1768 #else
1769         /* \\.\ is the device namespace in Windows, where every file
1770          * is a device node */
1771         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1772                 return false;
1773 #endif
1774
1775         return true;
1776 }
1777
1778 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1779 {
1780         int cur_files = td->files_index;
1781         char file_name[PATH_MAX];
1782         struct fio_file *f;
1783         int len = 0;
1784
1785         dprint(FD_FILE, "add file %s\n", fname);
1786
1787         if (td->o.directory)
1788                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1789                                         td->o.unique_filename);
1790
1791         sprintf(file_name + len, "%s", fname);
1792
1793         /* clean cloned siblings using existing files */
1794         if (numjob && is_already_allocated(file_name) &&
1795             !exists_and_not_regfile(fname))
1796                 return 0;
1797
1798         f = alloc_new_file(td);
1799
1800         if (td->files_size <= td->files_index) {
1801                 unsigned int new_size = td->o.nr_files + 1;
1802
1803                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1804
1805                 td->files = realloc(td->files, new_size * sizeof(f));
1806                 if (td->files == NULL) {
1807                         log_err("fio: realloc OOM\n");
1808                         assert(0);
1809                 }
1810                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1811                         td->file_locks = realloc(td->file_locks, new_size);
1812                         if (!td->file_locks) {
1813                                 log_err("fio: realloc OOM\n");
1814                                 assert(0);
1815                         }
1816                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1817                 }
1818                 td->files_size = new_size;
1819         }
1820         td->files[cur_files] = f;
1821         f->fileno = cur_files;
1822
1823         /*
1824          * init function, io engine may not be loaded yet
1825          */
1826         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1827                 f->real_file_size = -1ULL;
1828
1829         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1830                 f->file_name = strdup(file_name);
1831         else
1832                 f->file_name = smalloc_strdup(file_name);
1833
1834         /* can't handle smalloc failure from here */
1835         assert(f->file_name);
1836
1837         get_file_type(f);
1838
1839         switch (td->o.file_lock_mode) {
1840         case FILE_LOCK_NONE:
1841                 break;
1842         case FILE_LOCK_READWRITE:
1843                 f->rwlock = fio_rwlock_init();
1844                 break;
1845         case FILE_LOCK_EXCLUSIVE:
1846                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1847                 break;
1848         default:
1849                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1850                 assert(0);
1851         }
1852
1853         td->files_index++;
1854
1855         if (td->o.numjobs > 1)
1856                 set_already_allocated(file_name);
1857
1858         if (inc)
1859                 td->o.nr_files++;
1860
1861         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1862                                                         cur_files);
1863
1864         return cur_files;
1865 }
1866
1867 int add_file_exclusive(struct thread_data *td, const char *fname)
1868 {
1869         struct fio_file *f;
1870         unsigned int i;
1871
1872         for_each_file(td, f, i) {
1873                 if (!strcmp(f->file_name, fname))
1874                         return i;
1875         }
1876
1877         return add_file(td, fname, 0, 1);
1878 }
1879
1880 void get_file(struct fio_file *f)
1881 {
1882         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1883         assert(fio_file_open(f));
1884         f->references++;
1885 }
1886
1887 int put_file(struct thread_data *td, struct fio_file *f)
1888 {
1889         int f_ret = 0, ret = 0;
1890
1891         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1892
1893         if (!fio_file_open(f)) {
1894                 assert(f->fd == -1);
1895                 return 0;
1896         }
1897
1898         assert(f->references);
1899         if (--f->references)
1900                 return 0;
1901
1902         disk_util_dec(f->du);
1903
1904         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1905                 unlock_file_all(td, f);
1906
1907         if (should_fsync(td) && td->o.fsync_on_close) {
1908                 f_ret = fsync(f->fd);
1909                 if (f_ret < 0)
1910                         f_ret = errno;
1911         }
1912
1913         if (td->io_ops->close_file)
1914                 ret = td->io_ops->close_file(td, f);
1915
1916         if (!ret)
1917                 ret = f_ret;
1918
1919         td->nr_open_files--;
1920         fio_file_clear_closing(f);
1921         fio_file_clear_open(f);
1922         assert(f->fd == -1);
1923         return ret;
1924 }
1925
1926 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1927 {
1928         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1929                 return;
1930
1931         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1932                 if (ddir == DDIR_READ)
1933                         fio_rwlock_read(f->rwlock);
1934                 else
1935                         fio_rwlock_write(f->rwlock);
1936         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1937                 fio_sem_down(f->lock);
1938
1939         td->file_locks[f->fileno] = td->o.file_lock_mode;
1940 }
1941
1942 void unlock_file(struct thread_data *td, struct fio_file *f)
1943 {
1944         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1945                 return;
1946
1947         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1948                 fio_rwlock_unlock(f->rwlock);
1949         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1950                 fio_sem_up(f->lock);
1951
1952         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1953 }
1954
1955 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1956 {
1957         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1958                 return;
1959         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1960                 unlock_file(td, f);
1961 }
1962
1963 static bool recurse_dir(struct thread_data *td, const char *dirname)
1964 {
1965         struct dirent *dir;
1966         bool ret = false;
1967         DIR *D;
1968
1969         D = opendir(dirname);
1970         if (!D) {
1971                 char buf[FIO_VERROR_SIZE];
1972
1973                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1974                 td_verror(td, errno, buf);
1975                 return true;
1976         }
1977
1978         while ((dir = readdir(D)) != NULL) {
1979                 char full_path[PATH_MAX];
1980                 struct stat sb;
1981
1982                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1983                         continue;
1984
1985                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1986
1987                 if (lstat(full_path, &sb) == -1) {
1988                         if (errno != ENOENT) {
1989                                 td_verror(td, errno, "stat");
1990                                 ret = true;
1991                                 break;
1992                         }
1993                 }
1994
1995                 if (S_ISREG(sb.st_mode)) {
1996                         add_file(td, full_path, 0, 1);
1997                         continue;
1998                 }
1999                 if (!S_ISDIR(sb.st_mode))
2000                         continue;
2001
2002                 ret = recurse_dir(td, full_path);
2003                 if (ret)
2004                         break;
2005         }
2006
2007         closedir(D);
2008         return ret;
2009 }
2010
2011 int add_dir_files(struct thread_data *td, const char *path)
2012 {
2013         int ret = recurse_dir(td, path);
2014
2015         if (!ret)
2016                 log_info("fio: opendir added %d files\n", td->o.nr_files);
2017
2018         return ret;
2019 }
2020
2021 void dup_files(struct thread_data *td, struct thread_data *org)
2022 {
2023         struct fio_file *f;
2024         unsigned int i;
2025
2026         dprint(FD_FILE, "dup files: %d\n", org->files_index);
2027
2028         if (!org->files)
2029                 return;
2030
2031         td->files = calloc(org->files_index, sizeof(f));
2032
2033         if (td->o.file_lock_mode != FILE_LOCK_NONE)
2034                 td->file_locks = malloc(org->files_index);
2035
2036         assert(org->files_index >= org->o.nr_files);
2037         for_each_file(org, f, i) {
2038                 struct fio_file *__f;
2039
2040                 __f = alloc_new_file(td);
2041
2042                 if (f->file_name) {
2043                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2044                                 __f->file_name = strdup(f->file_name);
2045                         else
2046                                 __f->file_name = smalloc_strdup(f->file_name);
2047
2048                         /* can't handle smalloc failure from here */
2049                         assert(__f->file_name);
2050                         __f->filetype = f->filetype;
2051                 }
2052
2053                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2054                         __f->lock = f->lock;
2055                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2056                         __f->rwlock = f->rwlock;
2057
2058                 td->files[i] = __f;
2059         }
2060 }
2061
2062 /*
2063  * Returns the index that matches the filename, or -1 if not there
2064  */
2065 int get_fileno(struct thread_data *td, const char *fname)
2066 {
2067         struct fio_file *f;
2068         unsigned int i;
2069
2070         for_each_file(td, f, i)
2071                 if (!strcmp(f->file_name, fname))
2072                         return i;
2073
2074         return -1;
2075 }
2076
2077 /*
2078  * For log usage, where we add/open/close files automatically
2079  */
2080 void free_release_files(struct thread_data *td)
2081 {
2082         close_files(td);
2083         td->o.nr_files = 0;
2084         td->o.open_files = 0;
2085         td->files_index = 0;
2086 }
2087
2088 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2089 {
2090         int i;
2091
2092         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2093                 f->last_pos[i] = f->file_offset;
2094                 f->last_start[i] = -1ULL;
2095         }
2096
2097         if (fio_file_axmap(f))
2098                 axmap_reset(f->io_axmap);
2099         else if (fio_file_lfsr(f))
2100                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2101
2102         zbd_file_reset(td, f);
2103 }
2104
2105 bool fio_files_done(struct thread_data *td)
2106 {
2107         struct fio_file *f;
2108         unsigned int i;
2109
2110         for_each_file(td, f, i)
2111                 if (!fio_file_done(f))
2112                         return false;
2113
2114         return true;
2115 }
2116
2117 /* free memory used in initialization phase only */
2118 void filesetup_mem_free(void)
2119 {
2120         free_already_allocated();
2121 }
2122
2123 /*
2124  * This function is for platforms which support direct I/O but not O_DIRECT.
2125  */
2126 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2127 {
2128 #ifdef FIO_OS_DIRECTIO
2129         int ret = fio_set_odirect(f);
2130
2131         if (ret) {
2132                 td_verror(td, ret, "fio_set_directio");
2133 #if defined(__sun__)
2134                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2135                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2136                 } else {
2137                         log_err("fio: the file system does not seem to support direct IO\n");
2138                 }
2139 #else
2140                 log_err("fio: the file system does not seem to support direct IO\n");
2141 #endif
2142                 return -1;
2143         }
2144
2145         return 0;
2146 #else
2147         log_err("fio: direct IO is not supported on this host operating system\n");
2148         return -1;
2149 #endif
2150 }