filesetup.c

   1 #include <unistd.h>
   2 #include <fcntl.h>
   3 #include <string.h>
   4 #include <assert.h>
   5 #include <dirent.h>
   6 #include <libgen.h>
   7 #include <sys/stat.h>
   8
   9 #include "fio.h"
  10 #include "smalloc.h"
  11 #include "filehash.h"
  12 #include "options.h"
  13 #include "os/os.h"
  14 #include "hash.h"
  15 #include "lib/axmap.h"
  16 #include "rwlock.h"
  17 #include "zbd.h"
  18
  19 #ifdef CONFIG_LINUX_FALLOCATE
  20 #include <linux/falloc.h>
  21 #endif
  22
  23 static FLIST_HEAD(filename_list);
  24
  25 /*
  26  * List entry for filename_list
  27  */
  28 struct file_name {
  29         struct flist_head list;
  30         char *filename;
  31 };
  32
  33 static inline void clear_error(struct thread_data *td)
  34 {
  35         td->error = 0;
  36         td->verror[0] = '\0';
  37 }
  38
  39 static int native_fallocate(struct thread_data *td, struct fio_file *f)
  40 {
  41         bool success;
  42
  43         success = fio_fallocate(f, 0, f->real_file_size);
  44         dprint(FD_FILE, "native fallocate of file %s size %llu was "
  45                         "%ssuccessful\n", f->file_name,
  46                         (unsigned long long) f->real_file_size,
  47                         !success ? "un": "");
  48
  49         if (success)
  50                 return false;
  51
  52         if (errno == ENOSYS)
  53                 dprint(FD_FILE, "native fallocate is not implemented\n");
  54
  55         return true;
  56 }
  57
  58 static void fallocate_file(struct thread_data *td, struct fio_file *f)
  59 {
  60         if (td->o.fill_device)
  61                 return;
  62
  63         switch (td->o.fallocate_mode) {
  64         case FIO_FALLOCATE_NATIVE:
  65                 native_fallocate(td, f);
  66                 break;
  67         case FIO_FALLOCATE_NONE:
  68                 break;
  69 #ifdef CONFIG_POSIX_FALLOCATE
  70         case FIO_FALLOCATE_POSIX: {
  71                 int r;
  72
  73                 dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  74                                  f->file_name,
  75                                  (unsigned long long) f->real_file_size);
  76
  77                 r = posix_fallocate(f->fd, 0, f->real_file_size);
  78                 if (r > 0)
  79                         log_err("fio: posix_fallocate fails: %s\n", strerror(r));
  80                 break;
  81                 }
  82 #endif /* CONFIG_POSIX_FALLOCATE */
  83 #ifdef CONFIG_LINUX_FALLOCATE
  84         case FIO_FALLOCATE_KEEP_SIZE: {
  85                 int r;
  86
  87                 dprint(FD_FILE, "fallocate(FALLOC_FL_KEEP_SIZE) "
  88                                 "file %s size %llu\n", f->file_name,
  89                                 (unsigned long long) f->real_file_size);
  90
  91                 r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0, f->real_file_size);
  92                 if (r != 0)
  93                         td_verror(td, errno, "fallocate");
  94
  95                 break;
  96                 }
  97 #endif /* CONFIG_LINUX_FALLOCATE */
  98         case FIO_FALLOCATE_TRUNCATE: {
  99                 int r;
 100
 101                 dprint(FD_FILE, "ftruncate file %s size %llu\n",
 102                                 f->file_name,
 103                                 (unsigned long long) f->real_file_size);
 104                 r = ftruncate(f->fd, f->real_file_size);
 105                 if (r != 0)
 106                         td_verror(td, errno, "ftruncate");
 107
 108                 break;
 109         }
 110         default:
 111                 log_err("fio: unknown fallocate mode: %d\n", td->o.fallocate_mode);
 112                 assert(0);
 113         }
 114 }
 115
 116 /*
 117  * Leaves f->fd open on success, caller must close
 118  */
 119 static int extend_file(struct thread_data *td, struct fio_file *f)
 120 {
 121         int new_layout = 0, unlink_file = 0, flags;
 122         unsigned long long left;
 123         unsigned long long bs;
 124         char *b = NULL;
 125
 126         if (read_only) {
 127                 log_err("fio: refusing extend of file due to read-only\n");
 128                 return 0;
 129         }
 130
 131         /*
 132          * check if we need to lay the file out complete again. fio
 133          * does that for operations involving reads, or for writes
 134          * where overwrite is set
 135          */
 136         if (td_read(td) ||
 137            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
 138             (td_write(td) && td_ioengine_flagged(td, FIO_NOEXTEND)))
 139                 new_layout = 1;
 140         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
 141                 unlink_file = 1;
 142
 143         if (unlink_file || new_layout) {
 144                 int ret;
 145
 146                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
 147
 148                 ret = td_io_unlink_file(td, f);
 149                 if (ret != 0 && ret != ENOENT) {
 150                         td_verror(td, errno, "unlink");
 151                         return 1;
 152                 }
 153         }
 154
 155         flags = O_WRONLY;
 156         if (td->o.allow_create)
 157                 flags |= O_CREAT;
 158         if (new_layout)
 159                 flags |= O_TRUNC;
 160
 161 #ifdef WIN32
 162         flags |= _O_BINARY;
 163 #endif
 164
 165         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 166         f->fd = open(f->file_name, flags, 0644);
 167         if (f->fd < 0) {
 168                 int err = errno;
 169
 170                 if (err == ENOENT && !td->o.allow_create)
 171                         log_err("fio: file creation disallowed by "
 172                                         "allow_file_create=0\n");
 173                 else
 174                         td_verror(td, err, "open");
 175                 return 1;
 176         }
 177
 178         fallocate_file(td, f);
 179
 180         /*
 181          * If our jobs don't require regular files initially, we're done.
 182          */
 183         if (!new_layout)
 184                 goto done;
 185
 186         /*
 187          * The size will be -1ULL when fill_device is used, so don't truncate
 188          * or fallocate this file, just write it
 189          */
 190         if (!td->o.fill_device) {
 191                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 192                                         (unsigned long long) f->real_file_size);
 193                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 194                         if (errno != EFBIG) {
 195                                 td_verror(td, errno, "ftruncate");
 196                                 goto err;
 197                         }
 198                 }
 199         }
 200
 201         left = f->real_file_size;
 202         bs = td->o.max_bs[DDIR_WRITE];
 203         if (bs > left)
 204                 bs = left;
 205
 206         b = malloc(bs);
 207         if (!b) {
 208                 td_verror(td, errno, "malloc");
 209                 goto err;
 210         }
 211
 212         while (left && !td->terminate) {
 213                 ssize_t r;
 214
 215                 if (bs > left)
 216                         bs = left;
 217
 218                 fill_io_buffer(td, b, bs, bs);
 219
 220                 r = write(f->fd, b, bs);
 221
 222                 if (r > 0) {
 223                         left -= r;
 224                         continue;
 225                 } else {
 226                         if (r < 0) {
 227                                 int __e = errno;
 228
 229                                 if (__e == ENOSPC || __e == EDQUOT) {
 230                                         const char *__e_name;
 231                                         if (td->o.fill_device)
 232                                                 break;
 233                                         if (__e == ENOSPC)
 234                                                 __e_name = "ENOSPC";
 235                                         else
 236                                                 __e_name = "EDQUOT";
 237                                         log_info("fio: %s on laying out "
 238                                                  "file, stopping\n", __e_name);
 239                                 }
 240                                 td_verror(td, errno, "write");
 241                         } else
 242                                 td_verror(td, EIO, "write");
 243
 244                         goto err;
 245                 }
 246         }
 247
 248         if (td->terminate) {
 249                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 250                 td_io_unlink_file(td, f);
 251         } else if (td->o.create_fsync) {
 252                 if (fsync(f->fd) < 0) {
 253                         td_verror(td, errno, "fsync");
 254                         goto err;
 255                 }
 256         }
 257         if (td->o.fill_device && !td_write(td)) {
 258                 fio_file_clear_size_known(f);
 259                 if (td_io_get_file_size(td, f))
 260                         goto err;
 261                 if (f->io_size > f->real_file_size)
 262                         f->io_size = f->real_file_size;
 263         }
 264
 265         free(b);
 266 done:
 267         return 0;
 268 err:
 269         close(f->fd);
 270         f->fd = -1;
 271         if (b)
 272                 free(b);
 273         return 1;
 274 }
 275
 276 static bool pre_read_file(struct thread_data *td, struct fio_file *f)
 277 {
 278         int r, did_open = 0, old_runstate;
 279         unsigned long long left;
 280         unsigned long long bs;
 281         bool ret = true;
 282         char *b;
 283
 284         if (td_ioengine_flagged(td, FIO_PIPEIO) ||
 285             td_ioengine_flagged(td, FIO_NOIO))
 286                 return true;
 287
 288         if (f->filetype == FIO_TYPE_CHAR)
 289                 return true;
 290
 291         if (!fio_file_open(f)) {
 292                 if (td->io_ops->open_file(td, f)) {
 293                         log_err("fio: cannot pre-read, failed to open file\n");
 294                         return false;
 295                 }
 296                 did_open = 1;
 297         }
 298
 299         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 300
 301         left = f->io_size;
 302         bs = td->o.max_bs[DDIR_READ];
 303         if (bs > left)
 304                 bs = left;
 305
 306         b = calloc(1, bs);
 307         if (!b) {
 308                 td_verror(td, errno, "malloc");
 309                 ret = false;
 310                 goto error;
 311         }
 312
 313         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 314                 td_verror(td, errno, "lseek");
 315                 log_err("fio: failed to lseek pre-read file\n");
 316                 ret = false;
 317                 goto error;
 318         }
 319
 320         while (left && !td->terminate) {
 321                 if (bs > left)
 322                         bs = left;
 323
 324                 r = read(f->fd, b, bs);
 325
 326                 if (r == (int) bs) {
 327                         left -= bs;
 328                         continue;
 329                 } else {
 330                         td_verror(td, EIO, "pre_read");
 331                         break;
 332                 }
 333         }
 334
 335 error:
 336         td_restore_runstate(td, old_runstate);
 337
 338         if (did_open)
 339                 td->io_ops->close_file(td, f);
 340
 341         free(b);
 342         return ret;
 343 }
 344
 345 /*
 346  * Generic function to prepopulate regular file with data.
 347  * Useful if you want to make sure I/O engine has data to read.
 348  * Leaves f->fd open on success, caller must close.
 349  */
 350 int generic_prepopulate_file(struct thread_data *td, struct fio_file *f)
 351 {
 352         int flags;
 353         unsigned long long left, bs;
 354         char *b = NULL;
 355
 356         /* generic function for regular files only */
 357         assert(f->filetype == FIO_TYPE_FILE);
 358
 359         if (read_only) {
 360                 log_err("fio: refusing to write a file due to read-only\n");
 361                 return 0;
 362         }
 363
 364         flags = O_WRONLY;
 365         if (td->o.allow_create)
 366                 flags |= O_CREAT;
 367
 368 #ifdef WIN32
 369         flags |= _O_BINARY;
 370 #endif
 371
 372         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
 373         f->fd = open(f->file_name, flags, 0644);
 374         if (f->fd < 0) {
 375                 int err = errno;
 376
 377                 if (err == ENOENT && !td->o.allow_create)
 378                         log_err("fio: file creation disallowed by "
 379                                         "allow_file_create=0\n");
 380                 else
 381                         td_verror(td, err, "open");
 382                 return 1;
 383         }
 384
 385         left = f->real_file_size;
 386         bs = td->o.max_bs[DDIR_WRITE];
 387         if (bs > left)
 388                 bs = left;
 389
 390         b = malloc(bs);
 391         if (!b) {
 392                 td_verror(td, errno, "malloc");
 393                 goto err;
 394         }
 395
 396         while (left && !td->terminate) {
 397                 ssize_t r;
 398
 399                 if (bs > left)
 400                         bs = left;
 401
 402                 fill_io_buffer(td, b, bs, bs);
 403
 404                 r = write(f->fd, b, bs);
 405
 406                 if (r > 0) {
 407                         left -= r;
 408                 } else {
 409                         td_verror(td, errno, "write");
 410                         goto err;
 411                 }
 412         }
 413
 414         if (td->terminate) {
 415                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 416                 td_io_unlink_file(td, f);
 417         } else if (td->o.create_fsync) {
 418                 if (fsync(f->fd) < 0) {
 419                         td_verror(td, errno, "fsync");
 420                         goto err;
 421                 }
 422         }
 423
 424         free(b);
 425         return 0;
 426 err:
 427         close(f->fd);
 428         f->fd = -1;
 429         if (b)
 430                 free(b);
 431         return 1;
 432 }
 433
 434 unsigned long long get_rand_file_size(struct thread_data *td)
 435 {
 436         unsigned long long ret, sized;
 437         uint64_t frand_max;
 438         uint64_t r;
 439
 440         frand_max = rand_max(&td->file_size_state);
 441         r = __rand(&td->file_size_state);
 442         sized = td->o.file_size_high - td->o.file_size_low;
 443         ret = (unsigned long long) ((double) sized * (r / (frand_max + 1.0)));
 444         ret += td->o.file_size_low;
 445         ret -= (ret % td->o.rw_min_bs);
 446         return ret;
 447 }
 448
 449 static int file_size(struct thread_data *td, struct fio_file *f)
 450 {
 451         struct stat st;
 452
 453         if (stat(f->file_name, &st) == -1) {
 454                 td_verror(td, errno, "fstat");
 455                 return 1;
 456         }
 457
 458         f->real_file_size = st.st_size;
 459         return 0;
 460 }
 461
 462 static int bdev_size(struct thread_data *td, struct fio_file *f)
 463 {
 464         unsigned long long bytes = 0;
 465         int r;
 466
 467         if (td->io_ops->open_file(td, f)) {
 468                 log_err("fio: failed opening blockdev %s for size check\n",
 469                         f->file_name);
 470                 return 1;
 471         }
 472
 473         r = blockdev_size(f, &bytes);
 474         if (r) {
 475                 td_verror(td, r, "blockdev_size");
 476                 goto err;
 477         }
 478
 479         if (!bytes) {
 480                 log_err("%s: zero sized block device?\n", f->file_name);
 481                 goto err;
 482         }
 483
 484         f->real_file_size = bytes;
 485         td->io_ops->close_file(td, f);
 486         return 0;
 487 err:
 488         td->io_ops->close_file(td, f);
 489         return 1;
 490 }
 491
 492 static int char_size(struct thread_data *td, struct fio_file *f)
 493 {
 494 #ifdef FIO_HAVE_CHARDEV_SIZE
 495         unsigned long long bytes = 0;
 496         int r;
 497
 498         if (td->io_ops->open_file(td, f)) {
 499                 log_err("fio: failed opening chardev %s for size check\n",
 500                         f->file_name);
 501                 return 1;
 502         }
 503
 504         r = chardev_size(f, &bytes);
 505         if (r) {
 506                 td_verror(td, r, "chardev_size");
 507                 goto err;
 508         }
 509
 510         if (!bytes) {
 511                 log_err("%s: zero sized char device?\n", f->file_name);
 512                 goto err;
 513         }
 514
 515         f->real_file_size = bytes;
 516         td->io_ops->close_file(td, f);
 517         return 0;
 518 err:
 519         td->io_ops->close_file(td, f);
 520         return 1;
 521 #else
 522         f->real_file_size = -1ULL;
 523         return 0;
 524 #endif
 525 }
 526
 527 static int get_file_size(struct thread_data *td, struct fio_file *f)
 528 {
 529         int ret = 0;
 530
 531         if (fio_file_size_known(f))
 532                 return 0;
 533
 534         if (f->filetype == FIO_TYPE_FILE)
 535                 ret = file_size(td, f);
 536         else if (f->filetype == FIO_TYPE_BLOCK)
 537                 ret = bdev_size(td, f);
 538         else if (f->filetype == FIO_TYPE_CHAR)
 539                 ret = char_size(td, f);
 540         else {
 541                 f->real_file_size = -1;
 542                 log_info("%s: failed to get file size of %s\n", td->o.name,
 543                                         f->file_name);
 544                 return 1; /* avoid offset extends end error message */
 545         }
 546
 547         /*
 548          * Leave ->real_file_size with 0 since it could be expectation
 549          * of initial setup for regular files.
 550          */
 551         if (ret)
 552                 return ret;
 553
 554         /*
 555          * ->file_offset normally hasn't been initialized yet, so this
 556          * is basically always false unless ->real_file_size is -1, but
 557          * if ->real_file_size is -1 this message doesn't make sense.
 558          * As a result, this message is basically useless.
 559          */
 560         if (f->file_offset > f->real_file_size) {
 561                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 562                                         (unsigned long long) f->file_offset,
 563                                         (unsigned long long) f->real_file_size);
 564                 return 1;
 565         }
 566
 567         fio_file_set_size_known(f);
 568         return 0;
 569 }
 570
 571 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 572                                    unsigned long long off,
 573                                    unsigned long long len)
 574 {
 575         int errval = 0, ret = 0;
 576
 577 #ifdef CONFIG_ESX
 578         return 0;
 579 #endif
 580
 581         if (len == -1ULL)
 582                 len = f->io_size;
 583         if (off == -1ULL)
 584                 off = f->file_offset;
 585
 586         if (len == -1ULL || off == -1ULL)
 587                 return 0;
 588
 589         if (td->io_ops->invalidate) {
 590                 dprint(FD_IO, "invalidate %s cache %s\n", td->io_ops->name,
 591                         f->file_name);
 592                 ret = td->io_ops->invalidate(td, f);
 593                 if (ret < 0)
 594                         errval = -ret;
 595         } else if (td_ioengine_flagged(td, FIO_DISKLESSIO)) {
 596                 dprint(FD_IO, "invalidate not supported by ioengine %s\n",
 597                        td->io_ops->name);
 598         } else if (f->filetype == FIO_TYPE_FILE) {
 599                 dprint(FD_IO, "declare unneeded cache %s: %llu/%llu\n",
 600                         f->file_name, off, len);
 601                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 602                 if (ret)
 603                         errval = ret;
 604         } else if (f->filetype == FIO_TYPE_BLOCK) {
 605                 int retry_count = 0;
 606
 607                 dprint(FD_IO, "drop page cache %s\n", f->file_name);
 608                 ret = blockdev_invalidate_cache(f);
 609                 while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
 610                         /*
 611                          * Linux multipath devices reject ioctl while
 612                          * the maps are being updated. That window can
 613                          * last tens of milliseconds; we'll try up to
 614                          * a quarter of a second.
 615                          */
 616                         usleep(10000);
 617                         ret = blockdev_invalidate_cache(f);
 618                 }
 619                 if (ret < 0 && errno == EACCES && geteuid()) {
 620                         if (!fio_did_warn(FIO_WARN_ROOT_FLUSH)) {
 621                                 log_err("fio: only root may flush block "
 622                                         "devices. Cache flush bypassed!\n");
 623                         }
 624                 }
 625                 if (ret < 0)
 626                         errval = errno;
 627         } else if (f->filetype == FIO_TYPE_CHAR ||
 628                    f->filetype == FIO_TYPE_PIPE) {
 629                 dprint(FD_IO, "invalidate not supported %s\n", f->file_name);
 630         }
 631
 632         /*
 633          * Cache flushing isn't a fatal condition, and we know it will
 634          * happen on some platforms where we don't have the proper
 635          * function to flush eg block device caches. So just warn and
 636          * continue on our way.
 637          */
 638         if (errval)
 639                 log_info("fio: cache invalidation of %s failed: %s\n",
 640                          f->file_name, strerror(errval));
 641
 642         return 0;
 643
 644 }
 645
 646 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 647 {
 648         if (!fio_file_open(f))
 649                 return 0;
 650
 651         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 652 }
 653
 654 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 655 {
 656         int ret = 0;
 657
 658         dprint(FD_FILE, "fd close %s\n", f->file_name);
 659
 660         remove_file_hash(f);
 661
 662         if (close(f->fd) < 0)
 663                 ret = errno;
 664
 665         f->fd = -1;
 666
 667         if (f->shadow_fd != -1) {
 668                 close(f->shadow_fd);
 669                 f->shadow_fd = -1;
 670         }
 671
 672         f->engine_pos = 0;
 673         return ret;
 674 }
 675
 676 int file_lookup_open(struct fio_file *f, int flags)
 677 {
 678         struct fio_file *__f;
 679         int from_hash;
 680
 681         __f = lookup_file_hash(f->file_name);
 682         if (__f) {
 683                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 684                 f->lock = __f->lock;
 685                 from_hash = 1;
 686         } else {
 687                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 688                 from_hash = 0;
 689         }
 690
 691 #ifdef WIN32
 692         flags |= _O_BINARY;
 693 #endif
 694
 695         f->fd = open(f->file_name, flags, 0600);
 696         return from_hash;
 697 }
 698
 699 static int file_close_shadow_fds(struct thread_data *td)
 700 {
 701         struct fio_file *f;
 702         int num_closed = 0;
 703         unsigned int i;
 704
 705         for_each_file(td, f, i) {
 706                 if (f->shadow_fd == -1)
 707                         continue;
 708
 709                 close(f->shadow_fd);
 710                 f->shadow_fd = -1;
 711                 num_closed++;
 712         }
 713
 714         return num_closed;
 715 }
 716
 717 int generic_open_file(struct thread_data *td, struct fio_file *f)
 718 {
 719         int is_std = 0;
 720         int flags = 0;
 721         int from_hash = 0;
 722
 723         dprint(FD_FILE, "fd open %s\n", f->file_name);
 724
 725         if (!strcmp(f->file_name, "-")) {
 726                 if (td_rw(td)) {
 727                         log_err("fio: can't read/write to stdin/out\n");
 728                         return 1;
 729                 }
 730                 is_std = 1;
 731
 732                 /*
 733                  * move output logging to stderr, if we are writing to stdout
 734                  */
 735                 if (td_write(td))
 736                         f_out = stderr;
 737         }
 738
 739         if (td->o.odirect)
 740                 flags |= OS_O_DIRECT;
 741         flags |= td->o.sync_io;
 742         if (td->o.create_on_open && td->o.allow_create)
 743                 flags |= O_CREAT;
 744         if (f->filetype != FIO_TYPE_FILE)
 745                 flags |= FIO_O_NOATIME;
 746
 747 open_again:
 748         if (td_write(td)) {
 749                 if (!read_only)
 750                         flags |= O_RDWR;
 751
 752                 if (td->o.verify_only) {
 753                         flags &= ~O_RDWR;
 754                         flags |= O_RDONLY;
 755                 }
 756
 757                 if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
 758                         flags |= O_CREAT;
 759
 760                 if (is_std)
 761                         f->fd = dup(STDOUT_FILENO);
 762                 else
 763                         from_hash = file_lookup_open(f, flags);
 764         } else if (td_read(td)) {
 765                 if (td_ioengine_flagged(td, FIO_RO_NEEDS_RW_OPEN) && !read_only)
 766                         flags |= O_RDWR;
 767                 else
 768                         flags |= O_RDONLY;
 769
 770                 if (is_std)
 771                         f->fd = dup(STDIN_FILENO);
 772                 else
 773                         from_hash = file_lookup_open(f, flags);
 774         } else if (td_trim(td)) {
 775                 assert(!td_rw(td)); /* should have matched above */
 776                 if (!read_only)
 777                         flags |= O_RDWR;
 778                 from_hash = file_lookup_open(f, flags);
 779         }
 780
 781         if (f->fd == -1) {
 782                 char buf[FIO_VERROR_SIZE];
 783                 int __e = errno;
 784
 785                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 786                         flags &= ~FIO_O_NOATIME;
 787                         goto open_again;
 788                 }
 789                 if (__e == EMFILE && file_close_shadow_fds(td))
 790                         goto open_again;
 791
 792                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 793
 794                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 795                         log_err("fio: looks like your file system does not " \
 796                                 "support direct=1/buffered=0\n");
 797                 }
 798
 799                 td_verror(td, __e, buf);
 800                 return 1;
 801         }
 802
 803         if (!from_hash && f->fd != -1) {
 804                 if (add_file_hash(f)) {
 805                         int fio_unused ret;
 806
 807                         /*
 808                          * Stash away descriptor for later close. This is to
 809                          * work-around a "feature" on Linux, where a close of
 810                          * an fd that has been opened for write will trigger
 811                          * udev to call blkid to check partitions, fs id, etc.
 812                          * That pollutes the device cache, which can slow down
 813                          * unbuffered accesses.
 814                          */
 815                         if (f->shadow_fd == -1)
 816                                 f->shadow_fd = f->fd;
 817                         else {
 818                                 /*
 819                                  * OK to ignore, we haven't done anything
 820                                  * with it
 821                                  */
 822                                 ret = generic_close_file(td, f);
 823                         }
 824                         goto open_again;
 825                 }
 826         }
 827
 828         return 0;
 829 }
 830
 831 /*
 832  * This function i.e. get_file_size() is the default .get_file_size
 833  * implementation of majority of I/O engines.
 834  */
 835 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 836 {
 837         return get_file_size(td, f);
 838 }
 839
 840 /*
 841  * open/close all files, so that ->real_file_size gets set
 842  */
 843 static int get_file_sizes(struct thread_data *td)
 844 {
 845         struct fio_file *f;
 846         unsigned int i;
 847         int err = 0;
 848
 849         for_each_file(td, f, i) {
 850                 dprint(FD_FILE, "get file size for %p/%d/%s\n", f, i,
 851                                                                 f->file_name);
 852
 853                 if (td_io_get_file_size(td, f)) {
 854                         if (td->error != ENOENT) {
 855                                 log_err("%s\n", td->verror);
 856                                 err = 1;
 857                                 break;
 858                         }
 859                         clear_error(td);
 860                 }
 861
 862                 /*
 863                  * There are corner cases where we end up with -1 for
 864                  * ->real_file_size due to unsupported file type, etc.
 865                  * We then just set to size option value divided by number
 866                  * of files, similar to the way file ->io_size is set.
 867                  * stat(2) failure doesn't set ->real_file_size to -1.
 868                  */
 869                 if (f->real_file_size == -1ULL && td->o.size)
 870                         f->real_file_size = td->o.size / td->o.nr_files;
 871         }
 872
 873         return err;
 874 }
 875
 876 struct fio_mount {
 877         struct flist_head list;
 878         const char *base;
 879         char __base[256];
 880         unsigned int key;
 881 };
 882
 883 /*
 884  * Get free number of bytes for each file on each unique mount.
 885  */
 886 static unsigned long long get_fs_free_counts(struct thread_data *td)
 887 {
 888         struct flist_head *n, *tmp;
 889         unsigned long long ret = 0;
 890         struct fio_mount *fm;
 891         FLIST_HEAD(list);
 892         struct fio_file *f;
 893         unsigned int i;
 894
 895         for_each_file(td, f, i) {
 896                 struct stat sb;
 897                 char buf[256];
 898
 899                 if (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_CHAR) {
 900                         if (f->real_file_size != -1ULL)
 901                                 ret += f->real_file_size;
 902                         continue;
 903                 } else if (f->filetype != FIO_TYPE_FILE)
 904                         continue;
 905
 906                 snprintf(buf, FIO_ARRAY_SIZE(buf), "%s", f->file_name);
 907
 908                 if (stat(buf, &sb) < 0) {
 909                         if (errno != ENOENT)
 910                                 break;
 911                         strcpy(buf, ".");
 912                         if (stat(buf, &sb) < 0)
 913                                 break;
 914                 }
 915
 916                 fm = NULL;
 917                 flist_for_each(n, &list) {
 918                         fm = flist_entry(n, struct fio_mount, list);
 919                         if (fm->key == sb.st_dev)
 920                                 break;
 921
 922                         fm = NULL;
 923                 }
 924
 925                 if (fm)
 926                         continue;
 927
 928                 fm = calloc(1, sizeof(*fm));
 929                 snprintf(fm->__base, FIO_ARRAY_SIZE(fm->__base), "%s", buf);
 930                 fm->base = basename(fm->__base);
 931                 fm->key = sb.st_dev;
 932                 flist_add(&fm->list, &list);
 933         }
 934
 935         flist_for_each_safe(n, tmp, &list) {
 936                 unsigned long long sz;
 937
 938                 fm = flist_entry(n, struct fio_mount, list);
 939                 flist_del(&fm->list);
 940
 941                 sz = get_fs_free_size(fm->base);
 942                 if (sz && sz != -1ULL)
 943                         ret += sz;
 944
 945                 free(fm);
 946         }
 947
 948         return ret;
 949 }
 950
 951 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 952 {
 953         bool align = false;
 954         struct thread_options *o = &td->o;
 955         unsigned long long align_bs;
 956         unsigned long long offset;
 957         unsigned long long increment;
 958
 959         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 960                 return f->real_file_size;
 961
 962         if (o->offset_increment_percent) {
 963                 assert(!o->offset_increment);
 964                 increment = o->offset_increment_percent * f->real_file_size / 100;
 965                 align = true;
 966         } else
 967                 increment = o->offset_increment;
 968
 969         if (o->start_offset_percent > 0) {
 970                 /* calculate the raw offset */
 971                 offset = (f->real_file_size * o->start_offset_percent / 100) +
 972                         (td->subjob_number * increment);
 973
 974                 align = true;
 975         } else {
 976                 /* start_offset_percent not set */
 977                 offset = o->start_offset +
 978                                 td->subjob_number * increment;
 979         }
 980
 981         if (align) {
 982                 /*
 983                  * if offset_align is provided, use it
 984                  */
 985                 if (fio_option_is_set(o, start_offset_align)) {
 986                         align_bs = o->start_offset_align;
 987                 } else {
 988                         /* else take the minimum block size */
 989                         align_bs = td_min_bs(td);
 990                 }
 991
 992                 /*
 993                  * block align the offset at the next available boundary at
 994                  * ceiling(offset / align_bs) * align_bs
 995                  */
 996                 offset = (offset / align_bs + (offset % align_bs != 0)) * align_bs;
 997         }
 998
 999         return offset;
1000 }
1001
1002 /*
1003  * Find longest path component that exists and return its length
1004  */
1005 int longest_existing_path(char *path) {
1006         char buf[PATH_MAX];
1007         bool done;
1008         char *buf_pos;
1009         int offset;
1010 #ifdef WIN32
1011         DWORD dwAttr;
1012 #else
1013         struct stat sb;
1014 #endif
1015
1016         sprintf(buf, "%s", path);
1017         done = false;
1018         while (!done) {
1019                 buf_pos = strrchr(buf, FIO_OS_PATH_SEPARATOR);
1020                 if (!buf_pos) {
1021                         offset = 0;
1022                         break;
1023                 }
1024
1025                 *(buf_pos + 1) = '\0';
1026
1027 #ifdef WIN32
1028                 dwAttr = GetFileAttributesA(buf);
1029                 if (dwAttr != INVALID_FILE_ATTRIBUTES) {
1030                         done = true;
1031                 }
1032 #else
1033                 if (stat(buf, &sb) == 0)
1034                         done = true;
1035 #endif
1036                 if (done)
1037                         offset = buf_pos - buf;
1038                 else
1039                         *buf_pos = '\0';
1040         }
1041
1042         return offset;
1043 }
1044
1045 static bool create_work_dirs(struct thread_data *td, const char *fname)
1046 {
1047         char path[PATH_MAX];
1048         char *start, *end;
1049         int offset;
1050
1051         snprintf(path, PATH_MAX, "%s", fname);
1052         start = path;
1053
1054         offset = longest_existing_path(path);
1055         end = start + offset;
1056         while ((end = strchr(end, FIO_OS_PATH_SEPARATOR)) != NULL) {
1057                 if (end == start) {
1058                         end++;
1059                         continue;
1060                 }
1061                 *end = '\0';
1062                 errno = 0;
1063                 if (fio_mkdir(path, 0700) && errno != EEXIST) {
1064                         log_err("fio: failed to create dir (%s): %s\n",
1065                                 start, strerror(errno));
1066                         return false;
1067                 }
1068                 *end = FIO_OS_PATH_SEPARATOR;
1069                 end++;
1070         }
1071         td->flags |= TD_F_DIRS_CREATED;
1072         return true;
1073 }
1074
1075 /*
1076  * Open the files and setup files sizes, creating files if necessary.
1077  */
1078 int setup_files(struct thread_data *td)
1079 {
1080         unsigned long long total_size, extend_size;
1081         struct thread_options *o = &td->o;
1082         struct fio_file *f;
1083         unsigned int i, nr_fs_extra = 0;
1084         int err = 0, need_extend;
1085         int old_state;
1086         const unsigned long long bs = td_min_bs(td);
1087         uint64_t fs = 0;
1088
1089         dprint(FD_FILE, "setup files\n");
1090
1091         old_state = td_bump_runstate(td, TD_SETTING_UP);
1092
1093         for_each_file(td, f, i) {
1094                 if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1095                     strchr(f->file_name, FIO_OS_PATH_SEPARATOR) &&
1096                     !(td->flags & TD_F_DIRS_CREATED) &&
1097                     !create_work_dirs(td, f->file_name))
1098                         goto err_out;
1099         }
1100
1101         /*
1102          * Find out physical size of files or devices for this thread,
1103          * before we determine I/O size and range of our targets.
1104          * If ioengine defines a setup() method, it's responsible for
1105          * opening the files and setting f->real_file_size to indicate
1106          * the valid range for that file.
1107          */
1108         if (td->io_ops->setup)
1109                 err = td->io_ops->setup(td);
1110         else
1111                 err = get_file_sizes(td);
1112
1113         if (err)
1114                 goto err_out;
1115
1116         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1117                 err = zbd_init_files(td);
1118                 if (err)
1119                         goto err_out;
1120         }
1121         zbd_recalc_options_with_zone_granularity(td);
1122
1123         if (o->read_iolog_file)
1124                 goto done;
1125
1126         /*
1127          * check sizes. if the files/devices do not exist and the size
1128          * isn't passed to fio, abort.
1129          */
1130         total_size = 0;
1131         for_each_file(td, f, i) {
1132                 f->fileno = i;
1133                 if (f->real_file_size == -1ULL)
1134                         total_size = -1ULL;
1135                 else
1136                         total_size += f->real_file_size;
1137         }
1138
1139         if (o->fill_device)
1140                 td->fill_device_size = get_fs_free_counts(td);
1141
1142         /*
1143          * device/file sizes are zero and no size given, punt
1144          */
1145         if ((!total_size || total_size == -1ULL) && !o->size &&
1146             !td_ioengine_flagged(td, FIO_NOIO) && !o->fill_device &&
1147             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
1148                 log_err("%s: you need to specify size=\n", o->name);
1149                 td_verror(td, EINVAL, "total_file_size");
1150                 goto err_out;
1151         }
1152
1153         /*
1154          * Calculate per-file size and potential extra size for the
1155          * first files, if needed (i.e. if we don't have a fixed size).
1156          */
1157         if (!o->file_size_low && o->nr_files) {
1158                 uint64_t all_fs;
1159
1160                 fs = o->size / o->nr_files;
1161                 all_fs = fs * o->nr_files;
1162
1163                 if (all_fs < o->size)
1164                         nr_fs_extra = (o->size - all_fs) / bs;
1165         }
1166
1167         /*
1168          * now file sizes are known, so we can set ->io_size. if size= is
1169          * not given, ->io_size is just equal to ->real_file_size. if size
1170          * is given, ->io_size is size / nr_files.
1171          */
1172         extend_size = total_size = 0;
1173         need_extend = 0;
1174         for_each_file(td, f, i) {
1175                 f->file_offset = get_start_offset(td, f);
1176
1177                 /*
1178                  * Update ->io_size depending on options specified.
1179                  * ->file_size_low being 0 means filesize option isn't set.
1180                  * Non zero ->file_size_low equals ->file_size_high means
1181                  * filesize option is set in a fixed size format.
1182                  * Non zero ->file_size_low not equals ->file_size_high means
1183                  * filesize option is set in a range format.
1184                  */
1185                 if (!o->file_size_low) {
1186                         /*
1187                          * no file size or range given, file size is equal to
1188                          * total size divided by number of files. If the size
1189                          * doesn't divide nicely with the min blocksize,
1190                          * make the first files bigger.
1191                          */
1192                         f->io_size = fs;
1193                         if (nr_fs_extra) {
1194                                 nr_fs_extra--;
1195                                 f->io_size += bs;
1196                         }
1197
1198                         /*
1199                          * We normally don't come here for regular files, but
1200                          * if the result is 0 for a regular file, set it to the
1201                          * real file size. This could be size of the existing
1202                          * one if it already exists, but otherwise will be set
1203                          * to 0. A new file won't be created because
1204                          * ->io_size + ->file_offset equals ->real_file_size.
1205                          */
1206                         if (!f->io_size) {
1207                                 if (f->file_offset > f->real_file_size)
1208                                         goto err_offset;
1209                                 f->io_size = f->real_file_size - f->file_offset;
1210                                 if (!f->io_size)
1211                                         log_info("fio: file %s may be ignored\n",
1212                                                 f->file_name);
1213                         }
1214                 } else if (f->real_file_size < o->file_size_low ||
1215                            f->real_file_size > o->file_size_high) {
1216                         if (f->file_offset > o->file_size_low)
1217                                 goto err_offset;
1218                         /*
1219                          * file size given. if it's fixed, use that. if it's a
1220                          * range, generate a random size in-between.
1221                          */
1222                         if (o->file_size_low == o->file_size_high)
1223                                 f->io_size = o->file_size_low - f->file_offset;
1224                         else {
1225                                 f->io_size = get_rand_file_size(td)
1226                                                 - f->file_offset;
1227                         }
1228                 } else
1229                         f->io_size = f->real_file_size - f->file_offset;
1230
1231                 if (f->io_size == -1ULL)
1232                         total_size = -1ULL;
1233                 else {
1234                         uint64_t io_size;
1235
1236                         if (o->size_percent && o->size_percent != 100) {
1237                                 uint64_t file_size;
1238
1239                                 file_size = f->io_size + f->file_offset;
1240                                 f->io_size = (file_size *
1241                                               o->size_percent) / 100;
1242                                 if (f->io_size > (file_size - f->file_offset))
1243                                         f->io_size = file_size - f->file_offset;
1244
1245                                 f->io_size -= (f->io_size % td_min_bs(td));
1246                         }
1247
1248                         io_size = f->io_size;
1249                         if (o->io_size_percent && o->io_size_percent != 100) {
1250                                 io_size *= o->io_size_percent;
1251                                 io_size /= 100;
1252                         }
1253
1254                         total_size += io_size;
1255                 }
1256
1257                 if (f->filetype == FIO_TYPE_FILE &&
1258                     (f->io_size + f->file_offset) > f->real_file_size) {
1259                         if (!td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1260                             !o->create_on_open) {
1261                                 need_extend++;
1262                                 extend_size += (f->io_size + f->file_offset);
1263                                 fio_file_set_extend(f);
1264                         } else if (!td_ioengine_flagged(td, FIO_DISKLESSIO) ||
1265                                    (td_ioengine_flagged(td, FIO_DISKLESSIO) &&
1266                                     td_ioengine_flagged(td, FIO_FAKEIO)))
1267                                 f->real_file_size = f->io_size + f->file_offset;
1268                 }
1269         }
1270
1271         if (td->o.block_error_hist) {
1272                 int len;
1273
1274                 assert(td->o.nr_files == 1);    /* checked in fixup_options */
1275                 f = td->files[0];
1276                 len = f->io_size / td->o.bs[DDIR_TRIM];
1277                 if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
1278                         log_err("fio: cannot calculate block histogram with "
1279                                 "%d trim blocks, maximum %d\n",
1280                                 len, MAX_NR_BLOCK_INFOS);
1281                         td_verror(td, EINVAL, "block_error_hist");
1282                         goto err_out;
1283                 }
1284
1285                 td->ts.nr_block_infos = len;
1286                 for (i = 0; i < len; i++)
1287                         td->ts.block_infos[i] =
1288                                 BLOCK_INFO(0, BLOCK_STATE_UNINIT);
1289         } else
1290                 td->ts.nr_block_infos = 0;
1291
1292         if (!o->size || (total_size && o->size > total_size))
1293                 o->size = total_size;
1294
1295         if (o->size < td_min_bs(td)) {
1296                 log_err("fio: blocksize is larger than data set range\n");
1297                 goto err_out;
1298         }
1299
1300         /*
1301          * See if we need to extend some files, typically needed when our
1302          * target regular files don't exist yet, but our jobs require them
1303          * initially due to read I/Os.
1304          */
1305         if (need_extend) {
1306                 temp_stall_ts = 1;
1307                 if (output_format & FIO_OUTPUT_NORMAL) {
1308                         log_info("%s: Laying out IO file%s (%u file%s / %s%lluMiB)\n",
1309                                  o->name,
1310                                  need_extend > 1 ? "s" : "",
1311                                  need_extend,
1312                                  need_extend > 1 ? "s" : "",
1313                                  need_extend > 1 ? "total " : "",
1314                                  extend_size >> 20);
1315                 }
1316
1317                 for_each_file(td, f, i) {
1318                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
1319
1320                         if (!fio_file_extend(f))
1321                                 continue;
1322
1323                         assert(f->filetype == FIO_TYPE_FILE);
1324                         fio_file_clear_extend(f);
1325                         if (!o->fill_device) {
1326                                 old_len = f->real_file_size;
1327                                 extend_len = f->io_size + f->file_offset -
1328                                                 old_len;
1329                         }
1330                         f->real_file_size = (f->io_size + f->file_offset);
1331                         err = extend_file(td, f);
1332                         if (err)
1333                                 break;
1334
1335                         err = __file_invalidate_cache(td, f, old_len,
1336                                                                 extend_len);
1337
1338                         /*
1339                          * Shut up static checker
1340                          */
1341                         if (f->fd != -1)
1342                                 close(f->fd);
1343
1344                         f->fd = -1;
1345                         if (err)
1346                                 break;
1347                 }
1348                 temp_stall_ts = 0;
1349         }
1350
1351         if (err)
1352                 goto err_out;
1353
1354         /*
1355          * Prepopulate files with data. It might be expected to read some
1356          * "real" data instead of zero'ed files (if no writes to file occurred
1357          * prior to a read job). Engine has to provide a way to do that.
1358          */
1359         if (td->io_ops->prepopulate_file) {
1360                 temp_stall_ts = 1;
1361
1362                 for_each_file(td, f, i) {
1363                         if (output_format & FIO_OUTPUT_NORMAL) {
1364                                 log_info("%s: Prepopulating IO file (%s)\n",
1365                                                         o->name, f->file_name);
1366                         }
1367
1368                         err = td->io_ops->prepopulate_file(td, f);
1369                         if (err)
1370                                 break;
1371
1372                         err = __file_invalidate_cache(td, f, f->file_offset,
1373                                                                 f->io_size);
1374
1375                         /*
1376                          * Shut up static checker
1377                          */
1378                         if (f->fd != -1)
1379                                 close(f->fd);
1380
1381                         f->fd = -1;
1382                         if (err)
1383                                 break;
1384                 }
1385                 temp_stall_ts = 0;
1386         }
1387
1388         if (err)
1389                 goto err_out;
1390
1391         /*
1392          * iolog already set the total io size, if we read back
1393          * stored entries.
1394          */
1395         if (!o->read_iolog_file) {
1396                 if (o->io_size)
1397                         td->total_io_size = o->io_size * o->loops;
1398                 else
1399                         td->total_io_size = o->size * o->loops;
1400         }
1401
1402 done:
1403         if (td->o.zone_mode == ZONE_MODE_ZBD) {
1404                 err = zbd_setup_files(td);
1405                 if (err)
1406                         goto err_out;
1407         }
1408
1409         if (o->create_only)
1410                 td->done = 1;
1411
1412         td_restore_runstate(td, old_state);
1413
1414         if (td->o.dp_type != FIO_DP_NONE) {
1415                 err = dp_init(td);
1416                 if (err)
1417                         goto err_out;
1418         }
1419
1420         return 0;
1421
1422 err_offset:
1423         log_err("%s: you need to specify valid offset=\n", o->name);
1424 err_out:
1425         td_restore_runstate(td, old_state);
1426         return 1;
1427 }
1428
1429 bool pre_read_files(struct thread_data *td)
1430 {
1431         struct fio_file *f;
1432         unsigned int i;
1433
1434         dprint(FD_FILE, "pre_read files\n");
1435
1436         for_each_file(td, f, i) {
1437                 if (!pre_read_file(td, f))
1438                         return false;
1439         }
1440
1441         return true;
1442 }
1443
1444 static void __init_rand_distribution(struct thread_data *td, struct fio_file *f)
1445 {
1446         unsigned int range_size, seed;
1447         uint64_t nranges;
1448         uint64_t fsize;
1449
1450         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
1451         fsize = min(f->real_file_size, f->io_size);
1452
1453         nranges = (fsize + range_size - 1ULL) / range_size;
1454
1455         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number *
1456                 td->rand_seeds[FIO_RAND_BLOCK_OFF];
1457
1458         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
1459                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, td->o.random_center.u.f, seed);
1460         else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
1461                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, td->o.random_center.u.f, seed);
1462         else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
1463                 gauss_init(&f->gauss, nranges, td->o.gauss_dev.u.f, td->o.random_center.u.f, seed);
1464 }
1465
1466 static bool init_rand_distribution(struct thread_data *td)
1467 {
1468         struct fio_file *f;
1469         unsigned int i;
1470         int state;
1471
1472         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM ||
1473             td->o.random_distribution == FIO_RAND_DIST_ZONED ||
1474             td->o.random_distribution == FIO_RAND_DIST_ZONED_ABS)
1475                 return false;
1476
1477         state = td_bump_runstate(td, TD_SETTING_UP);
1478
1479         for_each_file(td, f, i)
1480                 __init_rand_distribution(td, f);
1481
1482         td_restore_runstate(td, state);
1483         return true;
1484 }
1485
1486 /*
1487  * Check if the number of blocks exceeds the randomness capability of
1488  * the selected generator. Tausworthe is 32-bit, the others are fully
1489  * 64-bit capable.
1490  */
1491 static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
1492                                  uint64_t blocks)
1493 {
1494         if (blocks <= FRAND32_MAX)
1495                 return 0;
1496         if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
1497                 return 0;
1498
1499         /*
1500          * If the user hasn't specified a random generator, switch
1501          * to tausworthe64 with informational warning. If the user did
1502          * specify one, just warn.
1503          */
1504         log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
1505                         f->file_name);
1506
1507         if (!fio_option_is_set(&td->o, random_generator)) {
1508                 log_info("fio: Switching to tausworthe64. Use the "
1509                          "random_generator= option to get rid of this "
1510                          "warning.\n");
1511                 td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
1512                 return 0;
1513         }
1514
1515         /*
1516          * Just make this information to avoid breaking scripts.
1517          */
1518         log_info("fio: Use the random_generator= option to switch to lfsr or "
1519                          "tausworthe64.\n");
1520         return 0;
1521 }
1522
1523 bool init_random_map(struct thread_data *td)
1524 {
1525         unsigned long long blocks;
1526         struct fio_file *f;
1527         unsigned int i;
1528
1529         if (init_rand_distribution(td))
1530                 return true;
1531         if (!td_random(td))
1532                 return true;
1533
1534         for_each_file(td, f, i) {
1535                 uint64_t fsize = min(f->real_file_size, f->io_size);
1536
1537                 if (td->o.zone_mode == ZONE_MODE_STRIDED)
1538                         fsize = td->o.zone_range;
1539
1540                 blocks = fsize / (unsigned long long) td->o.rw_min_bs;
1541
1542                 if (check_rand_gen_limits(td, f, blocks))
1543                         return false;
1544
1545                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1546                         uint64_t seed;
1547
1548                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1549
1550                         if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
1551                                 fio_file_set_lfsr(f);
1552                                 continue;
1553                         } else {
1554                                 log_err("fio: failed initializing LFSR\n");
1555                                 return false;
1556                         }
1557                 } else if (!td->o.norandommap) {
1558                         f->io_axmap = axmap_new(blocks);
1559                         if (f->io_axmap) {
1560                                 fio_file_set_axmap(f);
1561                                 continue;
1562                         }
1563                 } else if (td->o.norandommap)
1564                         continue;
1565
1566                 if (!td->o.softrandommap) {
1567                         log_err("fio: failed allocating random map. If running"
1568                                 " a large number of jobs, try the 'norandommap'"
1569                                 " option or set 'softrandommap'. Or give"
1570                                 " a larger --alloc-size to fio.\n");
1571                         return false;
1572                 }
1573
1574                 log_info("fio: file %s failed allocating random map. Running "
1575                          "job without.\n", f->file_name);
1576         }
1577
1578         return true;
1579 }
1580
1581 void close_files(struct thread_data *td)
1582 {
1583         struct fio_file *f;
1584         unsigned int i;
1585
1586         for_each_file(td, f, i) {
1587                 if (fio_file_open(f))
1588                         td_io_close_file(td, f);
1589         }
1590 }
1591
1592 void fio_file_free(struct fio_file *f)
1593 {
1594         if (fio_file_axmap(f))
1595                 axmap_free(f->io_axmap);
1596         if (f->ruhs_info)
1597                 sfree(f->ruhs_info);
1598         if (!fio_file_smalloc(f)) {
1599                 free(f->file_name);
1600                 free(f);
1601         } else {
1602                 sfree(f->file_name);
1603                 sfree(f);
1604         }
1605 }
1606
1607 void close_and_free_files(struct thread_data *td)
1608 {
1609         struct fio_file *f;
1610         unsigned int i;
1611
1612         dprint(FD_FILE, "close files\n");
1613
1614         for_each_file(td, f, i) {
1615                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1616                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1617                         td_io_unlink_file(td, f);
1618                 }
1619
1620                 if (fio_file_open(f))
1621                         td_io_close_file(td, f);
1622
1623                 remove_file_hash(f);
1624
1625                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1626                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1627                         td_io_unlink_file(td, f);
1628                 }
1629
1630                 zbd_close_file(f);
1631                 fdp_free_ruhs_info(f);
1632                 fio_file_free(f);
1633         }
1634
1635         td->o.filename = NULL;
1636         free(td->files);
1637         free(td->file_locks);
1638         td->files_index = 0;
1639         td->files = NULL;
1640         td->file_locks = NULL;
1641         td->o.file_lock_mode = FILE_LOCK_NONE;
1642         td->o.nr_files = 0;
1643 }
1644
1645 static void get_file_type(struct fio_file *f)
1646 {
1647         struct stat sb;
1648
1649         if (!strcmp(f->file_name, "-"))
1650                 f->filetype = FIO_TYPE_PIPE;
1651         else
1652                 f->filetype = FIO_TYPE_FILE;
1653
1654 #ifdef WIN32
1655         /* \\.\ is the device namespace in Windows, where every file is
1656          * a block device */
1657         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1658                 f->filetype = FIO_TYPE_BLOCK;
1659 #endif
1660
1661         if (!stat(f->file_name, &sb)) {
1662                 if (S_ISBLK(sb.st_mode))
1663                         f->filetype = FIO_TYPE_BLOCK;
1664                 else if (S_ISCHR(sb.st_mode))
1665                         f->filetype = FIO_TYPE_CHAR;
1666                 else if (S_ISFIFO(sb.st_mode))
1667                         f->filetype = FIO_TYPE_PIPE;
1668         }
1669 }
1670
1671 static bool __is_already_allocated(const char *fname, bool set)
1672 {
1673         struct flist_head *entry;
1674         bool ret;
1675
1676         ret = file_bloom_exists(fname, set);
1677         if (!ret)
1678                 return ret;
1679
1680         flist_for_each(entry, &filename_list) {
1681                 struct file_name *fn;
1682
1683                 fn = flist_entry(entry, struct file_name, list);
1684
1685                 if (!strcmp(fn->filename, fname))
1686                         return true;
1687         }
1688
1689         return false;
1690 }
1691
1692 static bool is_already_allocated(const char *fname)
1693 {
1694         bool ret;
1695
1696         fio_file_hash_lock();
1697         ret = __is_already_allocated(fname, false);
1698         fio_file_hash_unlock();
1699
1700         return ret;
1701 }
1702
1703 static void set_already_allocated(const char *fname)
1704 {
1705         struct file_name *fn;
1706
1707         fn = malloc(sizeof(struct file_name));
1708         fn->filename = strdup(fname);
1709
1710         fio_file_hash_lock();
1711         if (!__is_already_allocated(fname, true)) {
1712                 flist_add_tail(&fn->list, &filename_list);
1713                 fn = NULL;
1714         }
1715         fio_file_hash_unlock();
1716
1717         if (fn) {
1718                 free(fn->filename);
1719                 free(fn);
1720         }
1721 }
1722
1723 static void free_already_allocated(void)
1724 {
1725         struct flist_head *entry, *tmp;
1726         struct file_name *fn;
1727
1728         if (flist_empty(&filename_list))
1729                 return;
1730
1731         fio_file_hash_lock();
1732         flist_for_each_safe(entry, tmp, &filename_list) {
1733                 fn = flist_entry(entry, struct file_name, list);
1734                 free(fn->filename);
1735                 flist_del(&fn->list);
1736                 free(fn);
1737         }
1738
1739         fio_file_hash_unlock();
1740 }
1741
1742 static struct fio_file *alloc_new_file(struct thread_data *td)
1743 {
1744         struct fio_file *f;
1745
1746         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1747                 f = calloc(1, sizeof(*f));
1748         else
1749                 f = scalloc(1, sizeof(*f));
1750         if (!f) {
1751                 assert(0);
1752                 return NULL;
1753         }
1754
1755         f->fd = -1;
1756         f->shadow_fd = -1;
1757         fio_file_reset(td, f);
1758         if (!td_ioengine_flagged(td, FIO_NOFILEHASH))
1759                 fio_file_set_smalloc(f);
1760         return f;
1761 }
1762
1763 bool exists_and_not_regfile(const char *filename)
1764 {
1765         struct stat sb;
1766
1767         if (lstat(filename, &sb) == -1)
1768                 return false;
1769
1770 #ifndef WIN32 /* NOT Windows */
1771         if (S_ISREG(sb.st_mode))
1772                 return false;
1773 #else
1774         /* \\.\ is the device namespace in Windows, where every file
1775          * is a device node */
1776         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
1777                 return false;
1778 #endif
1779
1780         return true;
1781 }
1782
1783 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1784 {
1785         int cur_files = td->files_index;
1786         char file_name[PATH_MAX];
1787         struct fio_file *f;
1788         int len = 0;
1789
1790         dprint(FD_FILE, "add file %s\n", fname);
1791
1792         if (td->o.directory)
1793                 len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob,
1794                                         td->o.unique_filename);
1795
1796         sprintf(file_name + len, "%s", fname);
1797
1798         /* clean cloned siblings using existing files */
1799         if (numjob && is_already_allocated(file_name) &&
1800             !exists_and_not_regfile(fname))
1801                 return 0;
1802
1803         f = alloc_new_file(td);
1804
1805         if (td->files_size <= td->files_index) {
1806                 unsigned int new_size = td->o.nr_files + 1;
1807
1808                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1809
1810                 td->files = realloc(td->files, new_size * sizeof(f));
1811                 if (td->files == NULL) {
1812                         log_err("fio: realloc OOM\n");
1813                         assert(0);
1814                 }
1815                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1816                         td->file_locks = realloc(td->file_locks, new_size);
1817                         if (!td->file_locks) {
1818                                 log_err("fio: realloc OOM\n");
1819                                 assert(0);
1820                         }
1821                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1822                 }
1823                 td->files_size = new_size;
1824         }
1825         td->files[cur_files] = f;
1826         f->fileno = cur_files;
1827
1828         /*
1829          * init function, io engine may not be loaded yet
1830          */
1831         if (td->io_ops && td_ioengine_flagged(td, FIO_DISKLESSIO))
1832                 f->real_file_size = -1ULL;
1833
1834         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
1835                 f->file_name = strdup(file_name);
1836         else
1837                 f->file_name = smalloc_strdup(file_name);
1838
1839         /* can't handle smalloc failure from here */
1840         assert(f->file_name);
1841
1842         get_file_type(f);
1843
1844         switch (td->o.file_lock_mode) {
1845         case FILE_LOCK_NONE:
1846                 break;
1847         case FILE_LOCK_READWRITE:
1848                 f->rwlock = fio_rwlock_init();
1849                 break;
1850         case FILE_LOCK_EXCLUSIVE:
1851                 f->lock = fio_sem_init(FIO_SEM_UNLOCKED);
1852                 break;
1853         default:
1854                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1855                 assert(0);
1856         }
1857
1858         td->files_index++;
1859
1860         if (td->o.numjobs > 1)
1861                 set_already_allocated(file_name);
1862
1863         if (inc)
1864                 td->o.nr_files++;
1865
1866         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1867                                                         cur_files);
1868
1869         return cur_files;
1870 }
1871
1872 int add_file_exclusive(struct thread_data *td, const char *fname)
1873 {
1874         struct fio_file *f;
1875         unsigned int i;
1876
1877         for_each_file(td, f, i) {
1878                 if (!strcmp(f->file_name, fname))
1879                         return i;
1880         }
1881
1882         return add_file(td, fname, 0, 1);
1883 }
1884
1885 void get_file(struct fio_file *f)
1886 {
1887         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1888         assert(fio_file_open(f));
1889         f->references++;
1890 }
1891
1892 int put_file(struct thread_data *td, struct fio_file *f)
1893 {
1894         int f_ret = 0, ret = 0;
1895
1896         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1897
1898         if (!fio_file_open(f)) {
1899                 assert(f->fd == -1);
1900                 return 0;
1901         }
1902
1903         assert(f->references);
1904         if (--f->references)
1905                 return 0;
1906
1907         disk_util_dec(f->du);
1908
1909         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1910                 unlock_file_all(td, f);
1911
1912         if (should_fsync(td) && td->o.fsync_on_close) {
1913                 f_ret = fsync(f->fd);
1914                 if (f_ret < 0)
1915                         f_ret = errno;
1916         }
1917
1918         if (td->io_ops->close_file)
1919                 ret = td->io_ops->close_file(td, f);
1920
1921         if (!ret)
1922                 ret = f_ret;
1923
1924         td->nr_open_files--;
1925         fio_file_clear_closing(f);
1926         fio_file_clear_open(f);
1927         assert(f->fd == -1);
1928         return ret;
1929 }
1930
1931 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1932 {
1933         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1934                 return;
1935
1936         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1937                 if (ddir == DDIR_READ)
1938                         fio_rwlock_read(f->rwlock);
1939                 else
1940                         fio_rwlock_write(f->rwlock);
1941         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1942                 fio_sem_down(f->lock);
1943
1944         td->file_locks[f->fileno] = td->o.file_lock_mode;
1945 }
1946
1947 void unlock_file(struct thread_data *td, struct fio_file *f)
1948 {
1949         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1950                 return;
1951
1952         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1953                 fio_rwlock_unlock(f->rwlock);
1954         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1955                 fio_sem_up(f->lock);
1956
1957         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1958 }
1959
1960 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1961 {
1962         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1963                 return;
1964         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1965                 unlock_file(td, f);
1966 }
1967
1968 static bool recurse_dir(struct thread_data *td, const char *dirname)
1969 {
1970         struct dirent *dir;
1971         bool ret = false;
1972         DIR *D;
1973
1974         D = opendir(dirname);
1975         if (!D) {
1976                 char buf[FIO_VERROR_SIZE];
1977
1978                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1979                 td_verror(td, errno, buf);
1980                 return true;
1981         }
1982
1983         while ((dir = readdir(D)) != NULL) {
1984                 char full_path[PATH_MAX];
1985                 struct stat sb;
1986
1987                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1988                         continue;
1989
1990                 sprintf(full_path, "%s%c%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1991
1992                 if (lstat(full_path, &sb) == -1) {
1993                         if (errno != ENOENT) {
1994                                 td_verror(td, errno, "stat");
1995                                 ret = true;
1996                                 break;
1997                         }
1998                 }
1999
2000                 if (S_ISREG(sb.st_mode)) {
2001                         add_file(td, full_path, 0, 1);
2002                         continue;
2003                 }
2004                 if (!S_ISDIR(sb.st_mode))
2005                         continue;
2006
2007                 ret = recurse_dir(td, full_path);
2008                 if (ret)
2009                         break;
2010         }
2011
2012         closedir(D);
2013         return ret;
2014 }
2015
2016 int add_dir_files(struct thread_data *td, const char *path)
2017 {
2018         int ret = recurse_dir(td, path);
2019
2020         if (!ret)
2021                 log_info("fio: opendir added %d files\n", td->o.nr_files);
2022
2023         return ret;
2024 }
2025
2026 void dup_files(struct thread_data *td, struct thread_data *org)
2027 {
2028         struct fio_file *f;
2029         unsigned int i;
2030
2031         dprint(FD_FILE, "dup files: %d\n", org->files_index);
2032
2033         if (!org->files)
2034                 return;
2035
2036         td->files = calloc(org->files_index, sizeof(f));
2037
2038         if (td->o.file_lock_mode != FILE_LOCK_NONE)
2039                 td->file_locks = malloc(org->files_index);
2040
2041         assert(org->files_index >= org->o.nr_files);
2042         for_each_file(org, f, i) {
2043                 struct fio_file *__f;
2044
2045                 __f = alloc_new_file(td);
2046
2047                 if (f->file_name) {
2048                         if (td_ioengine_flagged(td, FIO_NOFILEHASH))
2049                                 __f->file_name = strdup(f->file_name);
2050                         else
2051                                 __f->file_name = smalloc_strdup(f->file_name);
2052
2053                         /* can't handle smalloc failure from here */
2054                         assert(__f->file_name);
2055                         __f->filetype = f->filetype;
2056                 }
2057
2058                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
2059                         __f->lock = f->lock;
2060                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
2061                         __f->rwlock = f->rwlock;
2062
2063                 td->files[i] = __f;
2064         }
2065 }
2066
2067 /*
2068  * Returns the index that matches the filename, or -1 if not there
2069  */
2070 int get_fileno(struct thread_data *td, const char *fname)
2071 {
2072         struct fio_file *f;
2073         unsigned int i;
2074
2075         for_each_file(td, f, i)
2076                 if (!strcmp(f->file_name, fname))
2077                         return i;
2078
2079         return -1;
2080 }
2081
2082 /*
2083  * For log usage, where we add/open/close files automatically
2084  */
2085 void free_release_files(struct thread_data *td)
2086 {
2087         close_files(td);
2088         td->o.nr_files = 0;
2089         td->o.open_files = 0;
2090         td->files_index = 0;
2091 }
2092
2093 void fio_file_reset(struct thread_data *td, struct fio_file *f)
2094 {
2095         int i;
2096
2097         for (i = 0; i < DDIR_RWDIR_CNT; i++) {
2098                 f->last_pos[i] = f->file_offset;
2099                 f->last_start[i] = -1ULL;
2100         }
2101
2102         if (fio_file_axmap(f))
2103                 axmap_reset(f->io_axmap);
2104         else if (fio_file_lfsr(f))
2105                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
2106
2107         zbd_file_reset(td, f);
2108 }
2109
2110 bool fio_files_done(struct thread_data *td)
2111 {
2112         struct fio_file *f;
2113         unsigned int i;
2114
2115         for_each_file(td, f, i)
2116                 if (!fio_file_done(f))
2117                         return false;
2118
2119         return true;
2120 }
2121
2122 /* free memory used in initialization phase only */
2123 void filesetup_mem_free(void)
2124 {
2125         free_already_allocated();
2126 }
2127
2128 /*
2129  * This function is for platforms which support direct I/O but not O_DIRECT.
2130  */
2131 int fio_set_directio(struct thread_data *td, struct fio_file *f)
2132 {
2133 #ifdef FIO_OS_DIRECTIO
2134         int ret = fio_set_odirect(f);
2135
2136         if (ret) {
2137                 td_verror(td, ret, "fio_set_directio");
2138 #if defined(__sun__)
2139                 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
2140                         log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
2141                 } else {
2142                         log_err("fio: the file system does not seem to support direct IO\n");
2143                 }
2144 #else
2145                 log_err("fio: the file system does not seem to support direct IO\n");
2146 #endif
2147                 return -1;
2148         }
2149
2150         return 0;
2151 #else
2152         log_err("fio: direct IO is not supported on this host operating system\n");
2153         return -1;
2154 #endif
2155 }