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