filesetup.c

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