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 %d unlink %d %s\n", new_layout, unlink_file, 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
 202         dprint(FD_FILE, "fill file %s, size %llu\n", f->file_name, (unsigned long long) f->real_file_size);
 203
 204         left = f->real_file_size;
 205         bs = td->o.max_bs[DDIR_WRITE];
 206         if (bs > left)
 207                 bs = left;
 208
 209         b = malloc(bs);
 210         if (!b) {
 211                 td_verror(td, errno, "malloc");
 212                 goto err;
 213         }
 214
 215         while (left && !td->terminate) {
 216                 ssize_t r;
 217
 218                 if (bs > left)
 219                         bs = left;
 220
 221                 fill_io_buffer(td, b, bs, bs);
 222
 223                 r = write(f->fd, b, bs);
 224
 225                 if (r > 0) {
 226                         left -= r;
 227                         continue;
 228                 } else {
 229                         if (r < 0) {
 230                                 int __e = errno;
 231
 232                                 if (__e == ENOSPC || __e == EDQUOT) {
 233                                         const char *__e_name;
 234                                         if (td->o.fill_device)
 235                                                 break;
 236                                         if (__e == ENOSPC)
 237                                                 __e_name = "ENOSPC";
 238                                         else
 239                                                 __e_name = "EDQUOT";
 240                                         log_info("fio: %s on laying out "
 241                                                  "file, stopping\n", __e_name);
 242                                 }
 243                                 td_verror(td, errno, "write");
 244                         } else
 245                                 td_verror(td, EIO, "write");
 246
 247                         goto err;
 248                 }
 249         }
 250
 251         if (td->terminate) {
 252                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 253                 td_io_unlink_file(td, f);
 254         } else if (td->o.create_fsync) {
 255                 if (fsync(f->fd) < 0) {
 256                         td_verror(td, errno, "fsync");
 257                         goto err;
 258                 }
 259         }
 260         if (td->o.fill_device && !td_write(td)) {
 261                 fio_file_clear_size_known(f);
 262                 if (td_io_get_file_size(td, f))
 263                         goto err;
 264                 if (f->io_size > f->real_file_size)
 265                         f->io_size = f->real_file_size;
 266         }
 267
 268         free(b);
 269 done:
 270         return 0;
 271 err:
 272         close(f->fd);
 273         f->fd = -1;
 274         if (b)
 275                 free(b);
 276         return 1;
 277 }
 278
 279 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 280 {
 281         int r, did_open = 0, old_runstate;
 282         unsigned long long left;
 283         unsigned long long bs;
 284         bool ret = true;
 285         char *b;
 286
 287         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 288             td_ioengine_flagged(td, FIO_NOIO))
 289                 return true;
 290
 291         if (f->filetype == FIO_TYPE_CHAR)
 292                 return true;
 293
 294         if (!fio_file_open(f)) {
 295                 if (td->io_ops->open_file(td, f)) {
 296                         log_err("fio: cannot pre-read, failed to open file\n");
 297                         return false;
 298                 }
 299                 did_open = 1;
 300         }
 301
 302         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 303
 304         left = f->io_size;
 305         bs = td->o.max_bs[DDIR_READ];
 306         if (bs > left)
 307                 bs = left;
 308
 309         b = malloc(bs);
 310         if (!b) {
 311                 td_verror(td, errno, "malloc");
 312                 ret = false;
 313                 goto error;
 314         }
 315         memset(b, 0, bs);
 316
 317         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 318                 td_verror(td, errno, "lseek");
 319                 log_err("fio: failed to lseek pre-read file\n");
 320                 ret = false;
 321                 goto error;
 322         }
 323
 324         while (left && !td->terminate) {
 325                 if (bs > left)
 326                         bs = left;
 327
 328                 r = read(f->fd, b, bs);
 329
 330                 if (r == (int) bs) {
 331                         left -= bs;
 332                         continue;
 333                 } else {
 334                         td_verror(td, EIO, "pre_read");
 335                         break;
 336                 }
 337         }
 338
 339 error:
 340         td_restore_runstate(td, old_runstate);
 341
 342         if (did_open)
 343                 td->io_ops->close_file(td, f);
 344
 345         free(b);
 346         return ret;
 347 }
 348
 349 /*
 350  * Generic function to prepopulate regular file with data.
 351  * Useful if you want to make sure I/O engine has data to read.
 352  * Leaves f->fd open on success, caller must close.
 353  */
 354 int generic_prepopulate_file(struct thread_data *td, struct fio_file *f)
 355 {
 356         int flags;
 357         unsigned long long left, bs;
 358         char *b = NULL;
 359
 360         /* generic function for regular files only */
 361         assert(f->filetype == FIO_TYPE_FILE);
 362
 363         if (read_only) {
 364                 log_err("fio: refusing to write a file due to read-only\n");
 365                 return 0;
 366         }
 367
 368         flags = O_WRONLY;
 369         if (td->o.allow_create)
 370                 flags |= O_CREAT;
 371
 372 #ifdef WIN32
 373         flags |= _O_BINARY;
 374 #endif
 375
 376         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 377         f->fd = open(f->file_name, flags, 0644);
 378         if (f->fd < 0) {
 379                 int err = errno;
 380
 381                 if (err == ENOENT && !td->o.allow_create)
 382                         log_err("fio: file creation disallowed by "
 383                                         "allow_file_create=0\n");
 384                 else
 385                         td_verror(td, err, "open");
 386                 return 1;
 387         }
 388
 389         left = f->real_file_size;
 390         bs = td->o.max_bs[DDIR_WRITE];
 391         if (bs > left)
 392                 bs = left;
 393
 394         b = malloc(bs);
 395         if (!b) {
 396                 td_verror(td, errno, "malloc");
 397                 goto err;
 398         }
 399
 400         while (left && !td->terminate) {
 401                 ssize_t r;
 402
 403                 if (bs > left)
 404                         bs = left;
 405
 406                 fill_io_buffer(td, b, bs, bs);
 407
 408                 r = write(f->fd, b, bs);
 409
 410                 if (r > 0) {
 411                         left -= r;
 412                 } else {
 413                         td_verror(td, errno, "write");
 414                         goto err;
 415                 }
 416         }
 417
 418         if (td->terminate) {
 419                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 420                 td_io_unlink_file(td, f);
 421         } else if (td->o.create_fsync) {
 422                 if (fsync(f->fd) < 0) {
 423                         td_verror(td, errno, "fsync");
 424                         goto err;
 425                 }
 426         }
 427
 428         free(b);
 429         return 0;
 430 err:
 431         close(f->fd);
 432         f->fd = -1;
 433         if (b)
 434                 free(b);
 435         return 1;
 436 }
 437
 438 unsigned long long get_rand_file_size(struct thread_data *td)
 439 {
 440         unsigned long long ret, sized;
 441         uint64_t frand_max;
 442         uint64_t r;
 443
 444         frand_max = rand_max(&td->file_size_state);
 445         r = __rand(&td->file_size_state);
 446         sized = td->o.file_size_high - td->o.file_size_low;
 447         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 448         ret += td->o.file_size_low;
 449         ret -= (ret % td->o.rw_min_bs);
 450         return ret;
 451 }
 452
 453 static int file_size(struct thread_data *td, struct fio_file *f)
 454 {
 455         struct stat st;
 456
 457         if (stat(f->file_name, &st) == -1) {
 458                 td_verror(td, errno, "fstat");
 459                 return 1;
 460         }
 461
 462         f->real_file_size = st.st_size;
 463         return 0;
 464 }
 465
 466 static int bdev_size(struct thread_data *td, struct fio_file *f)
 467 {
 468         unsigned long long bytes = 0;
 469         int r;
 470
 471         if (td->io_ops->open_file(td, f)) {
 472                 log_err("fio: failed opening blockdev %s for size check\n",
 473                         f->file_name);
 474                 return 1;
 475         }
 476
 477         r = blockdev_size(f, &bytes);
 478         if (r) {
 479                 td_verror(td, r, "blockdev_size");
 480                 goto err;
 481         }
 482
 483         if (!bytes) {
 484                 log_err("%s: zero sized block device?\n", f->file_name);
 485                 goto err;
 486         }
 487
 488         f->real_file_size = bytes;
 489         td->io_ops->close_file(td, f);
 490         return 0;
 491 err:
 492         td->io_ops->close_file(td, f);
 493         return 1;
 494 }
 495
 496 static int char_size(struct thread_data *td, struct fio_file *f)
 497 {
 498 #ifdef FIO_HAVE_CHARDEV_SIZE
 499         unsigned long long bytes = 0;
 500         int r;
 501
 502         if (td->io_ops->open_file(td, f)) {
 503                 log_err("fio: failed opening chardev %s for size check\n",
 504                         f->file_name);
 505                 return 1;
 506         }
 507
 508         r = chardev_size(f, &bytes);
 509         if (r) {
 510                 td_verror(td, r, "chardev_size");
 511                 goto err;
 512         }
 513
 514         if (!bytes) {
 515                 log_err("%s: zero sized char device?\n", f->file_name);
 516                 goto err;
 517         }
 518
 519         f->real_file_size = bytes;
 520         td->io_ops->close_file(td, f);
 521         return 0;
 522 err:
 523         td->io_ops->close_file(td, f);
 524         return 1;
 525 #else
 526         f->real_file_size = -1ULL;
 527         return 0;
 528 #endif
 529 }
 530
 531 static int get_file_size(struct thread_data *td, struct fio_file *f)
 532 {
 533         int ret = 0;
 534
 535         if (fio_file_size_known(f))
 536                 return 0;
 537
 538         if (f->filetype == FIO_TYPE_FILE)
 539                 ret = file_size(td, f);
 540         else if (f->filetype == FIO_TYPE_BLOCK)
 541                 ret = bdev_size(td, f);
 542         else if (f->filetype == FIO_TYPE_CHAR)
 543                 ret = char_size(td, f);
 544         else {
 545                 f->real_file_size = -1;
 546                 log_info("%s: failed to get file size of %s\n", td->o.name,
 547                                         f->file_name);
 548                 return 1; /* avoid offset extends end error message */
 549         }
 550
 551         /*
 552          * Leave ->real_file_size with 0 since it could be expectation
 553          * of initial setup for regular files.
 554          */
 555         if (ret)
 556                 return ret;
 557
 558         /*
 559          * ->file_offset normally hasn't been initialized yet, so this
 560          * is basically always false unless ->real_file_size is -1, but
 561          * if ->real_file_size is -1 this message doesn't make sense.
 562          * As a result, this message is basically useless.
 563          */
 564         if (f->file_offset > f->real_file_size) {
 565                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 566                                         (unsigned long long) f->file_offset,
 567                                         (unsigned long long) f->real_file_size);
 568                 return 1;
 569         }
 570
 571         fio_file_set_size_known(f);
 572         return 0;
 573 }
 574
 575 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 576                                    unsigned long long off,
 577                                    unsigned long long len)
 578 {
 579         int errval = 0, ret = 0;
 580
 581 #ifdef CONFIG_ESX
 582         return 0;
 583 #endif
 584
 585         if (len == -1ULL)
 586                 len = f->io_size;
 587         if (off == -1ULL)
 588                 off = f->file_offset;
 589
 590         if (len == -1ULL || off == -1ULL)
 591                 return 0;
 592
 593         if (td->io_ops->invalidate) {
 594                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 595                         f->file_name);
 596                 ret = td->io_ops->invalidate(td, f);
 597                 if (ret < 0)
 598                         errval = -ret;
 599         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 600                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
 601                        td->io_ops->name);
 602         } else if (f->filetype == FIO_TYPE_FILE) {
 603                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 604                         f->file_name, off, len);
 605                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 606                 if (ret)
 607                         errval = ret;
 608         } else if (f->filetype == FIO_TYPE_BLOCK) {
 609                 int retry_count = 0;
 610
 611                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 612                 ret = blockdev_invalidate_cache(f);
 613                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 614                         /*
 615                          * Linux multipath devices reject ioctl while
 616                          * the maps are being updated. That window can
 617                          * last tens of milliseconds; we'll try up to
 618                          * a quarter of a second.
 619                          */
 620                         usleep(10000);
 621                         ret = blockdev_invalidate_cache(f);
 622                 }
 623                 if (ret < 0 && errno == EACCES && geteuid()) {
 624                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 625                                 log_err("fio: only root may flush block "
 626                                         "devices. Cache flush bypassed!\n");
 627                         }
 628                 }
 629                 if (ret < 0)
 630                         errval = errno;
 631         } else if (f->filetype == FIO_TYPE_CHAR ||
 632                    f->filetype == FIO_TYPE_PIPE) {
 633                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 634         }
 635
 636         /*
 637          * Cache flushing isn't a fatal condition, and we know it will
 638          * happen on some platforms where we don't have the proper
 639          * function to flush eg block device caches. So just warn and
 640          * continue on our way.
 641          */
 642         if (errval)
 643                 log_info("fio: cache invalidation of %s failed: %s\n",
 644                          f->file_name, strerror(errval));
 645
 646         return 0;
 647
 648 }
 649
 650 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 651 {
 652         if (!fio_file_open(f))
 653                 return 0;
 654
 655         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 656 }
 657
 658 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 659 {
 660         int ret = 0;
 661
 662         dprint(FD_FILE, "fd close %s\n", f->file_name);
 663
 664         remove_file_hash(f);
 665
 666         if (close(f->fd) < 0)
 667                 ret = errno;
 668
 669         f->fd = -1;
 670
 671         if (f->shadow_fd != -1) {
 672                 close(f->shadow_fd);
 673                 f->shadow_fd = -1;
 674         }
 675
 676         f->engine_pos = 0;
 677         return ret;
 678 }
 679
 680 int file_lookup_open(struct fio_file *f, int flags)
 681 {
 682         struct fio_file *__f;
 683         int from_hash;
 684
 685         __f = lookup_file_hash(f->file_name);
 686         if (__f) {
 687                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 688                 f->lock = __f->lock;
 689                 from_hash = 1;
 690         } else {
 691                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 692                 from_hash = 0;
 693         }
 694
 695 #ifdef WIN32
 696         flags |= _O_BINARY;
 697 #endif
 698
 699         f->fd = open(f->file_name, flags, 0600);
 700         return from_hash;
 701 }
 702
 703 static int file_close_shadow_fds(struct thread_data *td)
 704 {
 705         struct fio_file *f;
 706         int num_closed = 0;
 707         unsigned int i;
 708
 709         for_each_file(td, f, i) {
 710                 if (f->shadow_fd == -1)
 711                         continue;
 712
 713                 close(f->shadow_fd);
 714                 f->shadow_fd = -1;
 715                 num_closed++;
 716         }
 717
 718         return num_closed;
 719 }
 720
 721 int generic_open_file(struct thread_data *td, struct fio_file *f)
 722 {
 723         int is_std = 0;
 724         int flags = 0;
 725         int from_hash = 0;
 726
 727         dprint(FD_FILE, "fd open %s\n", f->file_name);
 728
 729         if (!strcmp(f->file_name, "-")) {
 730                 if (td_rw(td)) {
 731                         log_err("fio: can't read/write to stdin/out\n");
 732                         return 1;
 733                 }
 734                 is_std = 1;
 735
 736                 /*
 737                  * move output logging to stderr, if we are writing to stdout
 738                  */
 739                 if (td_write(td))
 740                         f_out = stderr;
 741         }
 742
 743         if (td_trim(td))
 744                 goto skip_flags;
 745         if (td->o.odirect)
 746                 flags |= OS_O_DIRECT;
 747         if (td->o.oatomic) {
 748                 if (!FIO_O_ATOMIC) {
 749                         td_verror(td, EINVAL, "OS does not support atomic IO");
 750                         return 1;
 751                 }
 752                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 753         }
 754         flags |= td->o.sync_io;
 755         if (td->o.create_on_open && td->o.allow_create)
 756                 flags |= O_CREAT;
 757 skip_flags:
 758         if (f->filetype != FIO_TYPE_FILE)
 759                 flags |= FIO_O_NOATIME;
 760
 761 open_again:
 762         if (td_write(td)) {
 763                 if (!read_only)
 764                         flags |= O_RDWR;
 765
 766                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 767                         flags |= O_CREAT;
 768
 769                 if (is_std)
 770                         f->fd = dup(STDOUT_FILENO);
 771                 else
 772                         from_hash = file_lookup_open(f, flags);
 773         } else if (td_read(td)) {
 774                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 775                         flags |= O_RDWR;
 776                 else
 777                         flags |= O_RDONLY;
 778
 779                 if (is_std)
 780                         f->fd = dup(STDIN_FILENO);
 781                 else
 782                         from_hash = file_lookup_open(f, flags);
 783         } else if (td_trim(td)) {
 784                 assert(!td_rw(td)); /* should have matched above */
 785                 if (!read_only)
 786                         flags |= O_RDWR;
 787                 from_hash = file_lookup_open(f, flags);
 788         }
 789
 790         if (f->fd == -1) {
 791                 char buf[FIO_VERROR_SIZE];
 792                 int __e = errno;
 793
 794                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 795                         flags &= ~FIO_O_NOATIME;
 796                         goto open_again;
 797                 }
 798                 if (__e == EMFILE && file_close_shadow_fds(td))
 799                         goto open_again;
 800
 801                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 802
 803                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 804                         log_err("fio: looks like your file system does not " \
 805                                 "support direct=1/buffered=0\n");
 806                 }
 807
 808                 td_verror(td, __e, buf);
 809                 return 1;
 810         }
 811
 812         if (!from_hash && f->fd != -1) {
 813                 if (add_file_hash(f)) {
 814                         int fio_unused ret;
 815
 816                         /*
 817                          * Stash away descriptor for later close. This is to
 818                          * work-around a "feature" on Linux, where a close of
 819                          * an fd that has been opened for write will trigger
 820                          * udev to call blkid to check partitions, fs id, etc.
 821                          * That pollutes the device cache, which can slow down
 822                          * unbuffered accesses.
 823                          */
 824                         if (f->shadow_fd == -1)
 825                                 f->shadow_fd = f->fd;
 826                         else {
 827                                 /*
 828                                  * OK to ignore, we haven't done anything
 829                                  * with it
 830                                  */
 831                                 ret = generic_close_file(td, f);
 832                         }
 833                         goto open_again;
 834                 }
 835         }
 836
 837         return 0;
 838 }
 839
 840 /*
 841  * This function i.e. get_file_size() is the default .get_file_size
 842  * implementation of majority of I/O engines.
 843  */
 844 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 845 {
 846         return get_file_size(td, f);
 847 }
 848
 849 /*
 850  * open/close all files, so that ->real_file_size gets set
 851  */
 852 static int get_file_sizes(struct thread_data *td)
 853 {
 854         struct fio_file *f;
 855         unsigned int i;
 856         int err = 0;
 857
 858         for_each_file(td, f, i) {
 859                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 860                                                                 f->file_name);
 861
 862                 if (td_io_get_file_size(td, f)) {
 863                         if (td->error != ENOENT) {
 864                                 log_err("%s\n", td->verror);
 865                                 err = 1;
 866                                 break;
 867                         }
 868                         clear_error(td);
 869                 }
 870
 871                 /*
 872                  * There are corner cases where we end up with -1 for
 873                  * ->real_file_size due to unsupported file type, etc.
 874                  * We then just set to size option value divided by number
 875                  * of files, similar to the way file ->io_size is set.
 876                  * stat(2) failure doesn't set ->real_file_size to -1.
 877                  */
 878                 if (f->real_file_size == -1ULL && td->o.size)
 879                         f->real_file_size = td->o.size / td->o.nr_files;
 880         }
 881
 882         return err;
 883 }
 884
 885 struct fio_mount {
 886         struct flist_head list;
 887         const char *base;
 888         char __base[256];
 889         unsigned int key;
 890 };
 891
 892 /*
 893  * Get free number of bytes for each file on each unique mount.
 894  */
 895 static unsigned long long get_fs_free_counts(struct thread_data *td)
 896 {
 897         struct flist_head *n, *tmp;
 898         unsigned long long ret = 0;
 899         struct fio_mount *fm;
 900         FLIST_HEAD(list);
 901         struct fio_file *f;
 902         unsigned int i;
 903
 904         for_each_file(td, f, i) {
 905                 struct stat sb;
 906                 char buf[256];
 907
 908                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 909                         if (f->real_file_size != -1ULL)
 910                                 ret += f->real_file_size;
 911                         continue;
 912                 } else if (f->filetype != FIO_TYPE_FILE)
 913                         continue;
 914
 915                 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
 916
 917                 if (stat(buf, &sb) < 0) {
 918                         if (errno != ENOENT)
 919                                 break;
 920                         strcpy(buf, ".");
 921                         if (stat(buf, &sb) < 0)
 922                                 break;
 923                 }
 924
 925                 fm = NULL;
 926                 flist_for_each(n, &list) {
 927                         fm = flist_entry(n, struct fio_mount, list);
 928                         if (fm->key == sb.st_dev)
 929                                 break;
 930
 931                         fm = NULL;
 932                 }
 933
 934                 if (fm)
 935                         continue;
 936
 937                 fm = calloc(1, sizeof(*fm));
 938                 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
 939                 fm->base = basename(fm->__base);
 940                 fm->key = sb.st_dev;
 941                 flist_add(&fm->list, &list);
 942         }
 943
 944         flist_for_each_safe(n, tmp, &list) {
 945                 unsigned long long sz;
 946
 947                 fm = flist_entry(n, struct fio_mount, list);
 948                 flist_del(&fm->list);
 949
 950                 sz = get_fs_free_size(fm->base);
 951                 if (sz && sz != -1ULL)
 952                         ret += sz;
 953
 954                 free(fm);
 955         }
 956
 957         return ret;
 958 }
 959
 960 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 961 {
 962         bool align = false;
 963         struct thread_options *o = &td->o;
 964         unsigned long long align_bs;
 965         unsigned long long offset;
 966         unsigned long long increment;
 967
 968         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 969                 return f->real_file_size;
 970
 971         if (o->offset_increment_percent) {
 972                 assert(!o->offset_increment);
 973                 increment = o->offset_increment_percent * f->real_file_size / 100;
 974                 align = true;
 975         } else
 976                 increment = o->offset_increment;
 977
 978         if (o->start_offset_percent > 0) {
 979                 /* calculate the raw offset */
 980                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 981                         (td->subjob_number * increment);
 982
 983                 align = true;
 984         } else {
 985                 /* start_offset_percent not set */
 986                 offset = o->start_offset +
 987                                 td->subjob_number * increment;
 988         }
 989
 990         if (align) {
 991                 /*
 992                  * if offset_align is provided, use it
 993                  */
 994                 if (fio_option_is_set(o, start_offset_align)) {
 995                         align_bs = o->start_offset_align;
 996                 } else {
 997                         /* else take the minimum block size */
 998                         align_bs = td_min_bs(td);
 999                 }
1000
1001                 /*
1002                  * block align the offset at the next available boundary at
1003                  * ceiling(offset / align_bs) * align_bs
1004                  */
1005                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
1006         }
1007
1008         return offset;
1009 }
1010
1011 /*
1012  * Find longest path component that exists and return its length
1013  */
1014 int longest_existing_path(char *path) {
1015         char buf[PATH_MAX];
1016         bool done;
1017         char *buf_pos;
1018         int offset;
1019 #ifdef WIN32
1020         DWORD dwAttr;
1021 #else
1022         struct stat sb;
1023 #endif
1024
1025         sprintf(buf, "%s", path);
1026         done = false;
1027         while (!done) {
1028                 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1029                 if (!buf_pos) {
1030                         offset = 0;
1031                         break;
1032                 }
1033
1034                 *(buf_pos + 1) = '\0';
1035
1036 #ifdef WIN32
1037                 dwAttr = GetFileAttributesA(buf);
1038                 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1039                         done = true;
1040                 }
1041 #else
1042                 if (stat(buf, &sb) == 0)
1043                         done = true;
1044 #endif
1045                 if (done)
1046                         offset = buf_pos - buf;
1047                 else
1048                         *buf_pos = '\0';
1049         }
1050
1051         return offset;
1052 }
1053
1054 static bool create_work_dirs(struct thread_data *td, const char *fname)
1055 {
1056         char path[PATH_MAX];
1057         char *start, *end;
1058         int offset;
1059
1060         snprintf(path, PATH_MAX, "%s", fname);
1061         start = path;
1062
1063         offset = longest_existing_path(path);
1064         end = start + offset;
1065         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1066                 if (end == start) {
1067                         end++;
1068                         continue;
1069                 }
1070                 *end = '\0';
1071                 errno = 0;
1072                 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1073                         log_err("fio: failed to create dir (%s): %s\n",
1074                                 start, strerror(errno));
1075                         return false;
1076                 }
1077                 *end = FIO_OS_PATH_SEPARATOR;
1078                 end++;
1079         }
1080         td->flags |= TD_F_DIRS_CREATED;
1081         return true;
1082 }
1083
1084 int setup_shared_file(struct thread_data *td)
1085 {
1086         struct fio_file *f;
1087         uint64_t file_size;
1088         int err = 0;
1089
1090         if (td->o.nr_files > 1) {
1091                 log_err("fio: shared file setup called for multiple files\n");
1092                 return -1;
1093         }
1094
1095         get_file_sizes(td);
1096
1097         f = td->files[0];
1098
1099         if (f == NULL) {
1100                 log_err("fio: NULL shared file\n");
1101                 return -1;
1102         }
1103
1104         file_size = thread_number * td->o.size;
1105         dprint(FD_FILE, "shared setup %s real_file_size=%llu, desired=%llu\n",
1106                         f->file_name, (unsigned long long)f->real_file_size, (unsigned long long)file_size);
1107
1108         if (f->real_file_size < file_size) {
1109                 dprint(FD_FILE, "fio: extending shared file\n");
1110                 f->real_file_size = file_size;
1111                 err = extend_file(td, f);
1112                 if (!err)
1113                         err = __file_invalidate_cache(td, f, 0, f->real_file_size);
1114                 get_file_sizes(td);
1115                 dprint(FD_FILE, "shared setup new real_file_size=%llu\n",
1116                                 (unsigned long long)f->real_file_size);
1117         }
1118
1119         return err;
1120 }
1121
1122 /*
1123  * Open the files and setup files sizes, creating files if necessary.
1124  */
1125 int setup_files(struct thread_data *td)
1126 {
1127         unsigned long long total_size, extend_size;
1128         struct thread_options *o = &td->o;
1129         struct fio_file *f;
1130         unsigned int i, nr_fs_extra = 0;
1131         int err = 0, need_extend;
1132         int old_state;
1133         const unsigned long long bs = td_min_bs(td);
1134         uint64_t fs = 0;
1135
1136         dprint(FD_FILE, "setup files (thread_number=%d, subjob_number=%d)\n", td->thread_number, td->subjob_number);
1137
1138         old_state = td_bump_runstate(td, TD_SETTING_UP);
1139
1140         for_each_file(td, f, i) {
1141                 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1142                     strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1143                     !(td->flags & TD_F_DIRS_CREATED) &&
1144                     !create_work_dirs(td, f->file_name))
1145                         goto err_out;
1146         }
1147
1148         /*
1149          * Find out physical size of files or devices for this thread,
1150          * before we determine I/O size and range of our targets.
1151          * If ioengine defines a setup() method, it's responsible for
1152          * opening the files and setting f->real_file_size to indicate
1153          * the valid range for that file.
1154          */
1155         if (td->io_ops->setup)
1156                 err = td->io_ops->setup(td);
1157         else
1158                 err = get_file_sizes(td);
1159
1160         if (err)
1161                 goto err_out;
1162
1163         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1164                 err = zbd_init_files(td);
1165                 if (err)
1166                         goto err_out;
1167         }
1168         zbd_recalc_options_with_zone_granularity(td);
1169
1170         if (o->read_iolog_file)
1171                 goto done;
1172
1173         /*
1174          * check sizes. if the files/devices do not exist and the size
1175          * isn't passed to fio, abort.
1176          */
1177         total_size = 0;
1178         for_each_file(td, f, i) {
1179                 f->fileno = i;
1180                 if (f->real_file_size == -1ULL)
1181                         total_size = -1ULL;
1182                 else
1183                         total_size += f->real_file_size;
1184         }
1185
1186         if (o->fill_device)
1187                 td->fill_device_size = get_fs_free_counts(td);
1188
1189         /*
1190          * device/file sizes are zero and no size given, punt
1191          */
1192         if ((!total_size || total_size == -1ULL) && !o->size &&
1193             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1194             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1195                 log_err("%s: you need to specify size=\n", o->name);
1196                 td_verror(td, EINVAL, "total_file_size");
1197                 goto err_out;
1198         }
1199
1200         /*
1201          * Calculate per-file size and potential extra size for the
1202          * first files, if needed (i.e. if we don't have a fixed size).
1203          */
1204         if (!o->file_size_low && o->nr_files) {
1205                 uint64_t all_fs;
1206
1207                 fs = o->size / o->nr_files;
1208                 all_fs = fs * o->nr_files;
1209
1210                 if (all_fs < o->size)
1211                         nr_fs_extra = (o->size - all_fs) / bs;
1212         }
1213
1214         /*
1215          * now file sizes are known, so we can set ->io_size. if size= is
1216          * not given, ->io_size is just equal to ->real_file_size. if size
1217          * is given, ->io_size is size / nr_files.
1218          */
1219         extend_size = total_size = 0;
1220         need_extend = 0;
1221         for_each_file(td, f, i) {
1222                 f->file_offset = get_start_offset(td, f);
1223
1224                 /*
1225                  * Update ->io_size depending on options specified.
1226                  * ->file_size_low being 0 means filesize option isn't set.
1227                  * Non zero ->file_size_low equals ->file_size_high means
1228                  * filesize option is set in a fixed size format.
1229                  * Non zero ->file_size_low not equals ->file_size_high means
1230                  * filesize option is set in a range format.
1231                  */
1232                 if (!o->file_size_low) {
1233                         /*
1234                          * no file size or range given, file size is equal to
1235                          * total size divided by number of files. If the size
1236                          * doesn't divide nicely with the min blocksize,
1237                          * make the first files bigger.
1238                          */
1239                         f->io_size = fs;
1240                         if (nr_fs_extra) {
1241                                 nr_fs_extra--;
1242                                 f->io_size += bs;
1243                         }
1244
1245                         /*
1246                          * We normally don't come here for regular files, but
1247                          * if the result is 0 for a regular file, set it to the
1248                          * real file size. This could be size of the existing
1249                          * one if it already exists, but otherwise will be set
1250                          * to 0. A new file won't be created because
1251                          * ->io_size + ->file_offset equals ->real_file_size.
1252                          */
1253                         if (!f->io_size) {
1254                                 if (f->file_offset > f->real_file_size)
1255                                         goto err_offset;
1256                                 f->io_size = f->real_file_size - f->file_offset;
1257                                 if (!f->io_size)
1258                                         log_info("fio: file %s may be ignored\n",
1259                                                 f->file_name);
1260                         }
1261                 } else if (f->real_file_size < o->file_size_low ||
1262                            f->real_file_size > o->file_size_high) {
1263                         if (f->file_offset > o->file_size_low)
1264                                 goto err_offset;
1265                         /*
1266                          * file size given. if it's fixed, use that. if it's a
1267                          * range, generate a random size in-between.
1268                          */
1269                         if (o->file_size_low == o->file_size_high)
1270                                 f->io_size = o->file_size_low - f->file_offset;
1271                         else {
1272                                 f->io_size = get_rand_file_size(td)
1273                                                 - f->file_offset;
1274                         }
1275                 } else
1276                         f->io_size = f->real_file_size - f->file_offset;
1277
1278                 if (f->io_size == -1ULL)
1279                         total_size = -1ULL;
1280                 else {
1281                         uint64_t io_size;
1282
1283                         if (o->size_percent && o->size_percent != 100) {
1284                                 uint64_t file_size;
1285
1286                                 file_size = f->io_size + f->file_offset;
1287                                 f->io_size = (file_size *
1288                                               o->size_percent) / 100;
1289                                 if (f->io_size > (file_size - f->file_offset))
1290                                         f->io_size = file_size - f->file_offset;
1291
1292                                 f->io_size -= (f->io_size % td_min_bs(td));
1293                         }
1294
1295                         io_size = f->io_size;
1296                         if (o->io_size_percent && o->io_size_percent != 100) {
1297                                 io_size *= o->io_size_percent;
1298                                 io_size /= 100;
1299                         }
1300
1301                         total_size += io_size;
1302                 }
1303
1304                 if (f->filetype == FIO_TYPE_FILE &&
1305                     (f->io_size + f->file_offset) > f->real_file_size) {
1306                         if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1307                             !o->create_on_open) {
1308                                 need_extend++;
1309                                 extend_size += (f->io_size + f->file_offset);
1310                                 fio_file_set_extend(f);
1311                         } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1312                                    (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1313                                     td_ioengine_flagged(td, FIO_FAKEIO)))
1314                                 f->real_file_size = f->io_size + f->file_offset;
1315                 }
1316         }
1317
1318         if (td->o.block_error_hist) {
1319                 int len;
1320
1321                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1322                 f = td->files[0];
1323                 len = f->io_size / td->o.bs[DDIR_TRIM];
1324                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1325                         log_err("fio: cannot calculate block histogram with "
1326                                 "%d trim blocks, maximum %d\n",
1327                                 len, MAX_NR_BLOCK_INFOS);
1328                         td_verror(td, EINVAL, "block_error_hist");
1329                         goto err_out;
1330                 }
1331
1332                 td->ts.nr_block_infos = len;
1333                 for (i = 0; i < len; i++)
1334                         td->ts.block_infos[i] =
1335                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1336         } else
1337                 td->ts.nr_block_infos = 0;
1338
1339         if (!o->size || (total_size && o->size > total_size))
1340                 o->size = total_size;
1341
1342         if (o->size < td_min_bs(td)) {
1343                 log_err("fio: blocksize is larger than data set range\n");
1344                 goto err_out;
1345         }
1346
1347         /*
1348          * See if we need to extend some files, typically needed when our
1349          * target regular files don't exist yet, but our jobs require them
1350          * initially due to read I/Os.
1351          */
1352         if (need_extend) {
1353                 temp_stall_ts = 1;
1354                 if (output_format & FIO_OUTPUT_NORMAL) {
1355                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1356                                  o->name,
1357                                  need_extend > 1 ? "s" : "",
1358                                  need_extend,
1359                                  need_extend > 1 ? "s" : "",
1360                                  need_extend > 1 ? "total " : "",
1361                                  extend_size >> 20);
1362                 }
1363
1364                 for_each_file(td, f, i) {
1365                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1366
1367                         if (!fio_file_extend(f))
1368                                 continue;
1369
1370                         assert(f->filetype == FIO_TYPE_FILE);
1371                         fio_file_clear_extend(f);
1372                         if (!o->fill_device) {
1373                                 old_len = f->real_file_size;
1374                                 extend_len = f->io_size + f->file_offset -
1375                                                 old_len;
1376                         }
1377                         f->real_file_size = (f->io_size + f->file_offset);
1378                         err = extend_file(td, f);
1379                         if (err)
1380                                 break;
1381
1382                         err = __file_invalidate_cache(td, f, old_len,
1383                                                                 extend_len);
1384
1385                         /*
1386                          * Shut up static checker
1387                          */
1388                         if (f->fd != -1)
1389                                 close(f->fd);
1390
1391                         f->fd = -1;
1392                         if (err)
1393                                 break;
1394                 }
1395                 temp_stall_ts = 0;
1396         }
1397
1398         if (err)
1399                 goto err_out;
1400
1401         /*
1402          * Prepopulate files with data. It might be expected to read some
1403          * "real" data instead of zero'ed files (if no writes to file occurred
1404          * prior to a read job). Engine has to provide a way to do that.
1405          */
1406         if (td->io_ops->prepopulate_file) {
1407                 temp_stall_ts = 1;
1408
1409                 for_each_file(td, f, i) {
1410                         if (output_format & FIO_OUTPUT_NORMAL) {
1411                                 log_info("%s: Prepopulating IO file (%s)\n",
1412                                                         o->name, f->file_name);
1413                         }
1414
1415                         err = td->io_ops->prepopulate_file(td, f);
1416                         if (err)
1417                                 break;
1418
1419                         err = __file_invalidate_cache(td, f, f->file_offset,
1420                                                                 f->io_size);
1421
1422                         /*
1423                          * Shut up static checker
1424                          */
1425                         if (f->fd != -1)
1426                                 close(f->fd);
1427
1428                         f->fd = -1;
1429                         if (err)
1430                                 break;
1431                 }
1432                 temp_stall_ts = 0;
1433         }
1434
1435         if (err)
1436                 goto err_out;
1437
1438         /*
1439          * iolog already set the total io size, if we read back
1440          * stored entries.
1441          */
1442         if (!o->read_iolog_file) {
1443                 if (o->io_size)
1444                         td->total_io_size = o->io_size * o->loops;
1445                 else
1446                         td->total_io_size = o->size * o->loops;
1447         }
1448
1449 done:
1450         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1451                 err = zbd_setup_files(td);
1452                 if (err)
1453                         goto err_out;
1454         }
1455
1456         if (o->create_only)
1457                 td->done = 1;
1458
1459         td_restore_runstate(td, old_state);
1460
1461         return 0;
1462
1463 err_offset:
1464         log_err("%s: you need to specify valid offset=\n", o->name);
1465 err_out:
1466         td_restore_runstate(td, old_state);
1467         return 1;
1468 }
1469
1470 bool pre_read_files(struct thread_data *td)
1471 {
1472         struct fio_file *f;
1473         unsigned int i;
1474
1475         dprint(FD_FILE, "pre_read files\n");
1476
1477         for_each_file(td, f, i) {
1478                 if (!pre_read_file(td, f))
1479                         return false;
1480         }
1481
1482         return true;
1483 }
1484
1485 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1486 {
1487         unsigned int range_size, seed;
1488         uint64_t nranges;
1489         uint64_t fsize;
1490
1491         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1492         fsize = min(f->real_file_size, f->io_size);
1493
1494         nranges = (fsize + range_size - 1ULL) / range_size;
1495
1496         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
1497         if (!td->o.rand_repeatable)
1498                 seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1499
1500         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1501                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1502         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1503                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1504         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1505                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1506 }
1507
1508 static bool init_rand_distribution(struct thread_data *td)
1509 {
1510         struct fio_file *f;
1511         unsigned int i;
1512         int state;
1513
1514         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1515             td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1516             td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1517                 return false;
1518
1519         state = td_bump_runstate(td, TD_SETTING_UP);
1520
1521         for_each_file(td, f, i)
1522                 __init_rand_distribution(td, f);
1523
1524         td_restore_runstate(td, state);
1525         return true;
1526 }
1527
1528 /*
1529  * Check if the number of blocks exceeds the randomness capability of
1530  * the selected generator. Tausworthe is 32-bit, the others are fully
1531  * 64-bit capable.
1532  */
1533 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1534                                  uint64_t blocks)
1535 {
1536         if (blocks <= FRAND32_MAX)
1537                 return 0;
1538         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1539                 return 0;
1540
1541         /*
1542          * If the user hasn't specified a random generator, switch
1543          * to tausworthe64 with informational warning. If the user did
1544          * specify one, just warn.
1545          */
1546         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1547                         f->file_name);
1548
1549         if (!fio_option_is_set(&td->o, random_generator)) {
1550                 log_info("fio: Switching to tausworthe64. Use the "
1551                          "random_generator= option to get rid of this "
1552                          "warning.\n");
1553                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1554                 return 0;
1555         }
1556
1557         /*
1558          * Just make this information to avoid breaking scripts.
1559          */
1560         log_info("fio: Use the random_generator= option to switch to lfsr or "
1561                          "tausworthe64.\n");
1562         return 0;
1563 }
1564
1565 bool init_random_map(struct thread_data *td)
1566 {
1567         unsigned long long blocks;
1568         struct fio_file *f;
1569         unsigned int i;
1570
1571         if (init_rand_distribution(td))
1572                 return true;
1573         if (!td_random(td))
1574                 return true;
1575
1576         for_each_file(td, f, i) {
1577                 uint64_t fsize = min(f->real_file_size, f->io_size);
1578
1579                 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1580                         fsize = td->o.zone_range;
1581
1582                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1583
1584                 if (check_rand_gen_limits(td, f, blocks))
1585                         return false;
1586
1587                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1588                         uint64_t seed;
1589
1590                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1591
1592                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1593                                 fio_file_set_lfsr(f);
1594                                 continue;
1595                         } else {
1596                                 log_err("fio: failed initializing LFSR\n");
1597                                 return false;
1598                         }
1599                 } else if (!td->o.norandommap) {
1600                         f->io_axmap = axmap_new(blocks);
1601                         if (f->io_axmap) {
1602                                 fio_file_set_axmap(f);
1603                                 continue;
1604                         }
1605                 } else if (td->o.norandommap)
1606                         continue;
1607
1608                 if (!td->o.softrandommap) {
1609                         log_err("fio: failed allocating random map. If running"
1610                                 " a large number of jobs, try the 'norandommap'"
1611                                 " option or set 'softrandommap'. Or give"
1612                                 " a larger --alloc-size to fio.\n");
1613                         return false;
1614                 }
1615
1616                 log_info("fio: file %s failed allocating random map. Running "
1617                          "job without.\n", f->file_name);
1618         }
1619
1620         return true;
1621 }
1622
1623 void close_files(struct thread_data *td)
1624 {
1625         struct fio_file *f;
1626         unsigned int i;
1627
1628         for_each_file(td, f, i) {
1629                 if (fio_file_open(f))
1630                         td_io_close_file(td, f);
1631         }
1632 }
1633
1634 void fio_file_free(struct fio_file *f)
1635 {
1636         if (fio_file_axmap(f))
1637                 axmap_free(f->io_axmap);
1638         if (!fio_file_smalloc(f)) {
1639                 free(f->file_name);
1640                 free(f);
1641         } else {
1642                 sfree(f->file_name);
1643                 sfree(f);
1644         }
1645 }
1646
1647 void close_and_free_files(struct thread_data *td)
1648 {
1649         struct fio_file *f;
1650         unsigned int i;
1651
1652         dprint(FD_FILE, "close files\n");
1653
1654         for_each_file(td, f, i) {
1655                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1656                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1657                         td_io_unlink_file(td, f);
1658                 }
1659
1660                 if (fio_file_open(f))
1661                         td_io_close_file(td, f);
1662
1663                 remove_file_hash(f);
1664
1665                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1666                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1667                         td_io_unlink_file(td, f);
1668                 }
1669
1670                 zbd_close_file(f);
1671                 fio_file_free(f);
1672         }
1673
1674         td->o.filename = NULL;
1675         free(td->files);
1676         free(td->file_locks);
1677         td->files_index = 0;
1678         td->files = NULL;
1679         td->file_locks = NULL;
1680         td->o.file_lock_mode = FILE_LOCK_NONE;
1681         td->o.nr_files = 0;
1682 }
1683
1684 static void get_file_type(struct fio_file *f)
1685 {
1686         struct stat sb;
1687
1688         if (!strcmp(f->file_name, "-"))
1689                 f->filetype = FIO_TYPE_PIPE;
1690         else
1691                 f->filetype = FIO_TYPE_FILE;
1692
1693 #ifdef WIN32
1694         /* \\.\ is the device namespace in Windows, where every file is
1695          * a block device */
1696         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1697                 f->filetype = FIO_TYPE_BLOCK;
1698 #endif
1699
1700         if (!stat(f->file_name, &sb)) {
1701                 if (S_ISBLK(sb.st_mode))
1702                         f->filetype = FIO_TYPE_BLOCK;
1703                 else if (S_ISCHR(sb.st_mode))
1704                         f->filetype = FIO_TYPE_CHAR;
1705                 else if (S_ISFIFO(sb.st_mode))
1706                         f->filetype = FIO_TYPE_PIPE;
1707         }
1708 }
1709
1710 static bool __is_already_allocated(const char *fname, bool set)
1711 {
1712         struct flist_head *entry;
1713         bool ret;
1714
1715         ret = file_bloom_exists(fname, set);
1716         if (!ret)
1717                 return ret;
1718
1719         flist_for_each(entry, &filename_list) {
1720                 struct file_name *fn;
1721
1722                 fn = flist_entry(entry, struct file_name, list);
1723
1724                 if (!strcmp(fn->filename, fname))
1725                         return true;
1726         }
1727
1728         return false;
1729 }
1730
1731 static bool is_already_allocated(const char *fname)
1732 {
1733         bool ret;
1734
1735         fio_file_hash_lock();
1736         ret = __is_already_allocated(fname, false);
1737         fio_file_hash_unlock();
1738
1739         return ret;
1740 }
1741
1742 static void set_already_allocated(const char *fname)
1743 {
1744         struct file_name *fn;
1745
1746         fn = malloc(sizeof(struct file_name));
1747         fn->filename = strdup(fname);
1748
1749         fio_file_hash_lock();
1750         if (!__is_already_allocated(fname, true)) {
1751                 flist_add_tail(&fn->list, &filename_list);
1752                 fn = NULL;
1753         }
1754         fio_file_hash_unlock();
1755
1756         if (fn) {
1757                 free(fn->filename);
1758                 free(fn);
1759         }
1760 }
1761
1762 static void free_already_allocated(void)
1763 {
1764         struct flist_head *entry, *tmp;
1765         struct file_name *fn;
1766
1767         if (flist_empty(&filename_list))
1768                 return;
1769
1770         fio_file_hash_lock();
1771         flist_for_each_safe(entry, tmp, &filename_list) {
1772                 fn = flist_entry(entry, struct file_name, list);
1773                 free(fn->filename);
1774                 flist_del(&fn->list);
1775                 free(fn);
1776         }
1777
1778         fio_file_hash_unlock();
1779 }
1780
1781 static struct fio_file *alloc_new_file(struct thread_data *td)
1782 {
1783         struct fio_file *f;
1784
1785         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1786                 f = calloc(1, sizeof(*f));
1787         else
1788                 f = scalloc(1, sizeof(*f));
1789         if (!f) {
1790                 assert(0);
1791                 return NULL;
1792         }
1793
1794         f->fd = -1;
1795         f->shadow_fd = -1;
1796         fio_file_reset(td, f);
1797         if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1798                 fio_file_set_smalloc(f);
1799         return f;
1800 }
1801
1802 bool exists_and_not_regfile(const char *filename)
1803 {
1804         struct stat sb;
1805
1806         if (lstat(filename, &sb) == -1)
1807                 return false;
1808
1809 #ifndef WIN32 /* NOT Windows */
1810         if (S_ISREG(sb.st_mode))
1811                 return false;
1812 #else
1813         /* \\.\ is the device namespace in Windows, where every file
1814          * is a device node */
1815         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1816                 return false;
1817 #endif
1818
1819         return true;
1820 }
1821
1822 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1823 {
1824         int cur_files = td->files_index;
1825         char file_name[PATH_MAX];
1826         struct fio_file *f;
1827         int len = 0;
1828
1829         dprint(FD_FILE, "add file %s\n", fname);
1830
1831         if (td->o.directory)
1832                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1833                                         td->o.unique_filename);
1834
1835         sprintf(file_name + len, "%s", fname);
1836
1837         /* clean cloned siblings using existing files */
1838         if (numjob && is_already_allocated(file_name) &&
1839             !exists_and_not_regfile(fname))
1840                 return 0;
1841
1842         f = alloc_new_file(td);
1843
1844         if (td->files_size <= td->files_index) {
1845                 unsigned int new_size = td->o.nr_files + 1;
1846
1847                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1848
1849                 td->files = realloc(td->files, new_size * sizeof(f));
1850                 if (td->files == NULL) {
1851                         log_err("fio: realloc OOM\n");
1852                         assert(0);
1853                 }
1854                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1855                         td->file_locks = realloc(td->file_locks, new_size);
1856                         if (!td->file_locks) {
1857                                 log_err("fio: realloc OOM\n");
1858                                 assert(0);
1859                         }
1860                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1861                 }
1862                 td->files_size = new_size;
1863         }
1864         td->files[cur_files] = f;
1865         f->fileno = cur_files;
1866
1867         /*
1868          * init function, io engine may not be loaded yet
1869          */
1870         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1871                 f->real_file_size = -1ULL;
1872
1873         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1874                 f->file_name = strdup(file_name);
1875         else
1876                 f->file_name = smalloc_strdup(file_name);
1877
1878         /* can't handle smalloc failure from here */
1879         assert(f->file_name);
1880
1881         get_file_type(f);
1882
1883         switch (td->o.file_lock_mode) {
1884         case FILE_LOCK_NONE:
1885                 break;
1886         case FILE_LOCK_READWRITE:
1887                 f->rwlock = fio_rwlock_init();
1888                 break;
1889         case FILE_LOCK_EXCLUSIVE:
1890                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1891                 break;
1892         default:
1893                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1894                 assert(0);
1895         }
1896
1897         td->files_index++;
1898
1899         if (td->o.numjobs > 1)
1900                 set_already_allocated(file_name);
1901
1902         if (inc)
1903                 td->o.nr_files++;
1904
1905         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1906                                                         cur_files);
1907
1908         return cur_files;
1909 }
1910
1911 int add_file_exclusive(struct thread_data *td, const char *fname)
1912 {
1913         struct fio_file *f;
1914         unsigned int i;
1915
1916         for_each_file(td, f, i) {
1917                 if (!strcmp(f->file_name, fname))
1918                         return i;
1919         }
1920
1921         return add_file(td, fname, 0, 1);
1922 }
1923
1924 void get_file(struct fio_file *f)
1925 {
1926         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1927         assert(fio_file_open(f));
1928         f->references++;
1929 }
1930
1931 int put_file(struct thread_data *td, struct fio_file *f)
1932 {
1933         int f_ret = 0, ret = 0;
1934
1935         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1936
1937         if (!fio_file_open(f)) {
1938                 assert(f->fd == -1);
1939                 return 0;
1940         }
1941
1942         assert(f->references);
1943         if (--f->references)
1944                 return 0;
1945
1946         disk_util_dec(f->du);
1947
1948         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1949                 unlock_file_all(td, f);
1950
1951         if (should_fsync(td) && td->o.fsync_on_close) {
1952                 f_ret = fsync(f->fd);
1953                 if (f_ret < 0)
1954                         f_ret = errno;
1955         }
1956
1957         if (td->io_ops->close_file)
1958                 ret = td->io_ops->close_file(td, f);
1959
1960         if (!ret)
1961                 ret = f_ret;
1962
1963         td->nr_open_files--;
1964         fio_file_clear_closing(f);
1965         fio_file_clear_open(f);
1966         assert(f->fd == -1);
1967         return ret;
1968 }
1969
1970 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1971 {
1972         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1973                 return;
1974
1975         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1976                 if (ddir == DDIR_READ)
1977                         fio_rwlock_read(f->rwlock);
1978                 else
1979                         fio_rwlock_write(f->rwlock);
1980         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1981                 fio_sem_down(f->lock);
1982
1983         td->file_locks[f->fileno] = td->o.file_lock_mode;
1984 }
1985
1986 void unlock_file(struct thread_data *td, struct fio_file *f)
1987 {
1988         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1989                 return;
1990
1991         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1992                 fio_rwlock_unlock(f->rwlock);
1993         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1994                 fio_sem_up(f->lock);
1995
1996         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1997 }
1998
1999 void unlock_file_all(struct thread_data *td, struct fio_file *f)
2000 {
2001         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
2002                 return;
2003         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
2004                 unlock_file(td, f);
2005 }
2006
2007 static bool recurse_dir(struct thread_data *td, const char *dirname)
2008 {
2009         struct dirent *dir;
2010         bool ret = false;
2011         DIR *D;
2012
2013         D = opendir(dirname);
2014         if (!D) {
2015                 char buf[FIO_VERROR_SIZE];
2016
2017                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
2018                 td_verror(td, errno, buf);
2019                 return true;
2020         }
2021
2022         while ((dir = readdir(D)) != NULL) {
2023                 char full_path[PATH_MAX];
2024                 struct stat sb;
2025
2026                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
2027                         continue;
2028
2029                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
2030
2031                 if (lstat(full_path, &sb) == -1) {
2032                         if (errno != ENOENT) {
2033                                 td_verror(td, errno, "stat");
2034                                 ret = true;
2035                                 break;
2036                         }
2037                 }
2038
2039                 if (S_ISREG(sb.st_mode)) {
2040                         add_file(td, full_path, 0, 1);
2041                         continue;
2042                 }
2043                 if (!S_ISDIR(sb.st_mode))
2044                         continue;
2045
2046                 ret = recurse_dir(td, full_path);
2047                 if (ret)
2048                         break;
2049         }
2050
2051         closedir(D);
2052         return ret;
2053 }
2054
2055 int add_dir_files(struct thread_data *td, const char *path)
2056 {
2057         int ret = recurse_dir(td, path);
2058
2059         if (!ret)
2060                 log_info("fio: opendir added %d files\n", td->o.nr_files);
2061
2062         return ret;
2063 }
2064
2065 void dup_files(struct thread_data *td, struct thread_data *org)
2066 {
2067         struct fio_file *f;
2068         unsigned int i;
2069
2070         dprint(FD_FILE, "dup files: %d\n", org->files_index);
2071
2072         if (!org->files)
2073                 return;
2074
2075         td->files = calloc(org->files_index, sizeof(f));
2076
2077         if (td->o.file_lock_mode != FILE_LOCK_NONE)
2078                 td->file_locks = malloc(org->files_index);
2079
2080         assert(org->files_index >= org->o.nr_files);
2081         for_each_file(org, f, i) {
2082                 struct fio_file *__f;
2083
2084                 __f = alloc_new_file(td);
2085
2086                 if (f->file_name) {
2087                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2088                                 __f->file_name = strdup(f->file_name);
2089                         else
2090                                 __f->file_name = smalloc_strdup(f->file_name);
2091
2092                         /* can't handle smalloc failure from here */
2093                         assert(__f->file_name);
2094                         __f->filetype = f->filetype;
2095                 }
2096
2097                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2098                         __f->lock = f->lock;
2099                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2100                         __f->rwlock = f->rwlock;
2101
2102                 td->files[i] = __f;
2103         }
2104 }
2105
2106 /*
2107  * Returns the index that matches the filename, or -1 if not there
2108  */
2109 int get_fileno(struct thread_data *td, const char *fname)
2110 {
2111         struct fio_file *f;
2112         unsigned int i;
2113
2114         for_each_file(td, f, i)
2115                 if (!strcmp(f->file_name, fname))
2116                         return i;
2117
2118         return -1;
2119 }
2120
2121 /*
2122  * For log usage, where we add/open/close files automatically
2123  */
2124 void free_release_files(struct thread_data *td)
2125 {
2126         close_files(td);
2127         td->o.nr_files = 0;
2128         td->o.open_files = 0;
2129         td->files_index = 0;
2130 }
2131
2132 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2133 {
2134         int i;
2135
2136         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2137                 f->last_pos[i] = f->file_offset;
2138                 f->last_start[i] = -1ULL;
2139         }
2140
2141         if (fio_file_axmap(f))
2142                 axmap_reset(f->io_axmap);
2143         else if (fio_file_lfsr(f))
2144                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2145
2146         zbd_file_reset(td, f);
2147 }
2148
2149 bool fio_files_done(struct thread_data *td)
2150 {
2151         struct fio_file *f;
2152         unsigned int i;
2153
2154         for_each_file(td, f, i)
2155                 if (!fio_file_done(f))
2156                         return false;
2157
2158         return true;
2159 }
2160
2161 /* free memory used in initialization phase only */
2162 void filesetup_mem_free(void)
2163 {
2164         free_already_allocated();
2165 }
2166
2167 /*
2168  * This function is for platforms which support direct I/O but not O_DIRECT.
2169  */
2170 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2171 {
2172 #ifdef FIO_OS_DIRECTIO
2173         int ret = fio_set_odirect(f);
2174
2175         if (ret) {
2176                 td_verror(td, ret, "fio_set_directio");
2177 #if defined(__sun__)
2178                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2179                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2180                 } else {
2181                         log_err("fio: the file system does not seem to support direct IO\n");
2182                 }
2183 #else
2184                 log_err("fio: the file system does not seem to support direct IO\n");
2185 #endif
2186                 return -1;
2187         }
2188
2189         return 0;
2190 #else
2191         log_err("fio: direct IO is not supported on this host operating system\n");
2192         return -1;
2193 #endif
2194 }