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