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 #include <sys/mman.h>
   9 #include <sys/types.h>
  10
  11 #include "fio.h"
  12 #include "smalloc.h"
  13 #include "filehash.h"
  14 #include "os/os.h"
  15 #include "hash.h"
  16 #include "lib/axmap.h"
  17
  18 #ifdef CONFIG_LINUX_FALLOCATE
  19 #include <linux/falloc.h>
  20 #endif
  21
  22 static int root_warn;
  23
  24 static inline void clear_error(struct thread_data *td)
  25 {
  26         td->error = 0;
  27         td->verror[0] = '\0';
  28 }
  29
  30 /*
  31  * Leaves f->fd open on success, caller must close
  32  */
  33 static int extend_file(struct thread_data *td, struct fio_file *f)
  34 {
  35         int r, new_layout = 0, unlink_file = 0, flags;
  36         unsigned long long left;
  37         unsigned int bs;
  38         char *b;
  39
  40         if (read_only) {
  41                 log_err("fio: refusing extend of file due to read-only\n");
  42                 return 0;
  43         }
  44
  45         /*
  46          * check if we need to lay the file out complete again. fio
  47          * does that for operations involving reads, or for writes
  48          * where overwrite is set
  49          */
  50         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
  51             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
  52                 new_layout = 1;
  53         if (td_write(td) && !td->o.overwrite)
  54                 unlink_file = 1;
  55
  56         if (unlink_file || new_layout) {
  57                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  58                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
  59                         td_verror(td, errno, "unlink");
  60                         return 1;
  61                 }
  62         }
  63
  64         flags = O_WRONLY | O_CREAT;
  65         if (new_layout)
  66                 flags |= O_TRUNC;
  67
  68         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  69         f->fd = open(f->file_name, flags, 0644);
  70         if (f->fd < 0) {
  71                 td_verror(td, errno, "open");
  72                 return 1;
  73         }
  74
  75 #ifdef CONFIG_POSIX_FALLOCATE
  76         if (!td->o.fill_device) {
  77                 switch (td->o.fallocate_mode) {
  78                 case FIO_FALLOCATE_NONE:
  79                         break;
  80                 case FIO_FALLOCATE_POSIX:
  81                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  82                                  f->file_name,
  83                                  (unsigned long long) f->real_file_size);
  84
  85                         r = posix_fallocate(f->fd, 0, f->real_file_size);
  86                         if (r > 0) {
  87                                 log_err("fio: posix_fallocate fails: %s\n",
  88                                                 strerror(r));
  89                         }
  90                         break;
  91 #ifdef CONFIG_LINUX_FALLOCATE
  92                 case FIO_FALLOCATE_KEEP_SIZE:
  93                         dprint(FD_FILE,
  94                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
  95                                 "file %s size %llu\n", f->file_name,
  96                                 (unsigned long long) f->real_file_size);
  97
  98                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
  99                                         f->real_file_size);
 100                         if (r != 0)
 101                                 td_verror(td, errno, "fallocate");
 102
 103                         break;
 104 #endif /* CONFIG_LINUX_FALLOCATE */
 105                 default:
 106                         log_err("fio: unknown fallocate mode: %d\n",
 107                                 td->o.fallocate_mode);
 108                         assert(0);
 109                 }
 110         }
 111 #endif /* CONFIG_POSIX_FALLOCATE */
 112
 113         if (!new_layout)
 114                 goto done;
 115
 116         /*
 117          * The size will be -1ULL when fill_device is used, so don't truncate
 118          * or fallocate this file, just write it
 119          */
 120         if (!td->o.fill_device) {
 121                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 122                                         (unsigned long long) f->real_file_size);
 123                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 124                         td_verror(td, errno, "ftruncate");
 125                         goto err;
 126                 }
 127         }
 128
 129         b = malloc(td->o.max_bs[DDIR_WRITE]);
 130
 131         left = f->real_file_size;
 132         while (left && !td->terminate) {
 133                 bs = td->o.max_bs[DDIR_WRITE];
 134                 if (bs > left)
 135                         bs = left;
 136
 137                 fill_io_buffer(td, b, bs, bs);
 138
 139                 r = write(f->fd, b, bs);
 140
 141                 if (r > 0) {
 142                         left -= r;
 143                         continue;
 144                 } else {
 145                         if (r < 0) {
 146                                 int __e = errno;
 147
 148                                 if (__e == ENOSPC) {
 149                                         if (td->o.fill_device)
 150                                                 break;
 151                                         log_info("fio: ENOSPC on laying out "
 152                                                  "file, stopping\n");
 153                                         break;
 154                                 }
 155                                 td_verror(td, errno, "write");
 156                         } else
 157                                 td_verror(td, EIO, "write");
 158
 159                         break;
 160                 }
 161         }
 162
 163         if (td->terminate) {
 164                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 165                 unlink(f->file_name);
 166         } else if (td->o.create_fsync) {
 167                 if (fsync(f->fd) < 0) {
 168                         td_verror(td, errno, "fsync");
 169                         goto err;
 170                 }
 171         }
 172         if (td->o.fill_device && !td_write(td)) {
 173                 fio_file_clear_size_known(f);
 174                 if (td_io_get_file_size(td, f))
 175                         goto err;
 176                 if (f->io_size > f->real_file_size)
 177                         f->io_size = f->real_file_size;
 178         }
 179
 180         free(b);
 181 done:
 182         return 0;
 183 err:
 184         close(f->fd);
 185         f->fd = -1;
 186         return 1;
 187 }
 188
 189 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 190 {
 191         int r, did_open = 0, old_runstate;
 192         unsigned long long left;
 193         unsigned int bs;
 194         char *b;
 195
 196         if (td->io_ops->flags & FIO_PIPEIO)
 197                 return 0;
 198
 199         if (!fio_file_open(f)) {
 200                 if (td->io_ops->open_file(td, f)) {
 201                         log_err("fio: cannot pre-read, failed to open file\n");
 202                         return 1;
 203                 }
 204                 did_open = 1;
 205         }
 206
 207         old_runstate = td->runstate;
 208         td_set_runstate(td, TD_PRE_READING);
 209
 210         bs = td->o.max_bs[DDIR_READ];
 211         b = malloc(bs);
 212         memset(b, 0, bs);
 213
 214         lseek(f->fd, f->file_offset, SEEK_SET);
 215         left = f->io_size;
 216
 217         while (left && !td->terminate) {
 218                 if (bs > left)
 219                         bs = left;
 220
 221                 r = read(f->fd, b, bs);
 222
 223                 if (r == (int) bs) {
 224                         left -= bs;
 225                         continue;
 226                 } else {
 227                         td_verror(td, EIO, "pre_read");
 228                         break;
 229                 }
 230         }
 231
 232         td_set_runstate(td, old_runstate);
 233
 234         if (did_open)
 235                 td->io_ops->close_file(td, f);
 236         free(b);
 237         return 0;
 238 }
 239
 240 static unsigned long long get_rand_file_size(struct thread_data *td)
 241 {
 242         unsigned long long ret, sized;
 243         unsigned long r;
 244
 245         if (td->o.use_os_rand) {
 246                 r = os_random_long(&td->file_size_state);
 247                 sized = td->o.file_size_high - td->o.file_size_low;
 248                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 249         } else {
 250                 r = __rand(&td->__file_size_state);
 251                 sized = td->o.file_size_high - td->o.file_size_low;
 252                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 253         }
 254
 255         ret += td->o.file_size_low;
 256         ret -= (ret % td->o.rw_min_bs);
 257         return ret;
 258 }
 259
 260 static int file_size(struct thread_data *td, struct fio_file *f)
 261 {
 262         struct stat st;
 263
 264         if (stat(f->file_name, &st) == -1) {
 265                 td_verror(td, errno, "fstat");
 266                 return 1;
 267         }
 268
 269         f->real_file_size = st.st_size;
 270         return 0;
 271 }
 272
 273 static int bdev_size(struct thread_data *td, struct fio_file *f)
 274 {
 275         unsigned long long bytes = 0;
 276         int r;
 277
 278         if (td->io_ops->open_file(td, f)) {
 279                 log_err("fio: failed opening blockdev %s for size check\n",
 280                         f->file_name);
 281                 return 1;
 282         }
 283
 284         r = blockdev_size(f, &bytes);
 285         if (r) {
 286                 td_verror(td, r, "blockdev_size");
 287                 goto err;
 288         }
 289
 290         if (!bytes) {
 291                 log_err("%s: zero sized block device?\n", f->file_name);
 292                 goto err;
 293         }
 294
 295         f->real_file_size = bytes;
 296         td->io_ops->close_file(td, f);
 297         return 0;
 298 err:
 299         td->io_ops->close_file(td, f);
 300         return 1;
 301 }
 302
 303 static int char_size(struct thread_data *td, struct fio_file *f)
 304 {
 305 #ifdef FIO_HAVE_CHARDEV_SIZE
 306         unsigned long long bytes = 0;
 307         int r;
 308
 309         if (td->io_ops->open_file(td, f)) {
 310                 log_err("fio: failed opening blockdev %s for size check\n",
 311                         f->file_name);
 312                 return 1;
 313         }
 314
 315         r = chardev_size(f, &bytes);
 316         if (r) {
 317                 td_verror(td, r, "chardev_size");
 318                 goto err;
 319         }
 320
 321         if (!bytes) {
 322                 log_err("%s: zero sized char device?\n", f->file_name);
 323                 goto err;
 324         }
 325
 326         f->real_file_size = bytes;
 327         td->io_ops->close_file(td, f);
 328         return 0;
 329 err:
 330         td->io_ops->close_file(td, f);
 331         return 1;
 332 #else
 333         f->real_file_size = -1ULL;
 334         return 0;
 335 #endif
 336 }
 337
 338 static int get_file_size(struct thread_data *td, struct fio_file *f)
 339 {
 340         int ret = 0;
 341
 342         if (fio_file_size_known(f))
 343                 return 0;
 344
 345         if (f->filetype == FIO_TYPE_FILE)
 346                 ret = file_size(td, f);
 347         else if (f->filetype == FIO_TYPE_BD)
 348                 ret = bdev_size(td, f);
 349         else if (f->filetype == FIO_TYPE_CHAR)
 350                 ret = char_size(td, f);
 351         else
 352                 f->real_file_size = -1;
 353
 354         if (ret)
 355                 return ret;
 356
 357         if (f->file_offset > f->real_file_size) {
 358                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 359                                         (unsigned long long) f->file_offset,
 360                                         (unsigned long long) f->real_file_size);
 361                 return 1;
 362         }
 363
 364         fio_file_set_size_known(f);
 365         return 0;
 366 }
 367
 368 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 369                                    unsigned long long off,
 370                                    unsigned long long len)
 371 {
 372         int ret = 0;
 373
 374         if (len == -1ULL)
 375                 len = f->io_size;
 376         if (off == -1ULL)
 377                 off = f->file_offset;
 378
 379         if (len == -1ULL || off == -1ULL)
 380                 return 0;
 381
 382         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 383                                                                 len);
 384
 385         /*
 386          * FIXME: add blockdev flushing too
 387          */
 388         if (f->mmap_ptr) {
 389                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 390 #ifdef FIO_MADV_FREE
 391                 if (f->filetype == FIO_TYPE_BD)
 392                         (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 393 #endif
 394         } else if (f->filetype == FIO_TYPE_FILE) {
 395                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 396         } else if (f->filetype == FIO_TYPE_BD) {
 397                 ret = blockdev_invalidate_cache(f);
 398                 if (ret < 0 && errno == EACCES && geteuid()) {
 399                         if (!root_warn) {
 400                                 log_err("fio: only root may flush block "
 401                                         "devices. Cache flush bypassed!\n");
 402                                 root_warn = 1;
 403                         }
 404                         ret = 0;
 405                 }
 406         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 407                 ret = 0;
 408
 409         if (ret < 0) {
 410                 td_verror(td, errno, "invalidate_cache");
 411                 return 1;
 412         } else if (ret > 0) {
 413                 td_verror(td, ret, "invalidate_cache");
 414                 return 1;
 415         }
 416
 417         return ret;
 418
 419 }
 420
 421 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 422 {
 423         if (!fio_file_open(f))
 424                 return 0;
 425
 426         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 427 }
 428
 429 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 430 {
 431         int ret = 0;
 432
 433         dprint(FD_FILE, "fd close %s\n", f->file_name);
 434
 435         remove_file_hash(f);
 436
 437         if (close(f->fd) < 0)
 438                 ret = errno;
 439
 440         f->fd = -1;
 441
 442         if (f->shadow_fd != -1) {
 443                 close(f->shadow_fd);
 444                 f->shadow_fd = -1;
 445         }
 446
 447         f->engine_data = 0;
 448         return ret;
 449 }
 450
 451 int file_lookup_open(struct fio_file *f, int flags)
 452 {
 453         struct fio_file *__f;
 454         int from_hash;
 455
 456         __f = lookup_file_hash(f->file_name);
 457         if (__f) {
 458                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 459                 /*
 460                  * racy, need the __f->lock locked
 461                  */
 462                 f->lock = __f->lock;
 463                 from_hash = 1;
 464         } else {
 465                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 466                 from_hash = 0;
 467         }
 468
 469         f->fd = open(f->file_name, flags, 0600);
 470         return from_hash;
 471 }
 472
 473 static int file_close_shadow_fds(struct thread_data *td)
 474 {
 475         struct fio_file *f;
 476         int num_closed = 0;
 477         unsigned int i;
 478
 479         for_each_file(td, f, i) {
 480                 if (f->shadow_fd == -1)
 481                         continue;
 482
 483                 close(f->shadow_fd);
 484                 f->shadow_fd = -1;
 485                 num_closed++;
 486         }
 487
 488         return num_closed;
 489 }
 490
 491 int generic_open_file(struct thread_data *td, struct fio_file *f)
 492 {
 493         int is_std = 0;
 494         int flags = 0;
 495         int from_hash = 0;
 496
 497         dprint(FD_FILE, "fd open %s\n", f->file_name);
 498
 499         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 500                 log_err("fio: trim only applies to block device\n");
 501                 return 1;
 502         }
 503
 504         if (!strcmp(f->file_name, "-")) {
 505                 if (td_rw(td)) {
 506                         log_err("fio: can't read/write to stdin/out\n");
 507                         return 1;
 508                 }
 509                 is_std = 1;
 510
 511                 /*
 512                  * move output logging to stderr, if we are writing to stdout
 513                  */
 514                 if (td_write(td))
 515                         f_out = stderr;
 516         }
 517
 518         if (td_trim(td))
 519                 goto skip_flags;
 520         if (td->o.odirect)
 521                 flags |= OS_O_DIRECT;
 522         if (td->o.oatomic) {
 523                 if (!FIO_O_ATOMIC) {
 524                         td_verror(td, EINVAL, "OS does not support atomic IO");
 525                         return 1;
 526                 }
 527                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 528         }
 529         if (td->o.sync_io)
 530                 flags |= O_SYNC;
 531         if (td->o.create_on_open)
 532                 flags |= O_CREAT;
 533 skip_flags:
 534         if (f->filetype != FIO_TYPE_FILE)
 535                 flags |= FIO_O_NOATIME;
 536
 537 open_again:
 538         if (td_write(td)) {
 539                 if (!read_only)
 540                         flags |= O_RDWR;
 541
 542                 if (f->filetype == FIO_TYPE_FILE)
 543                         flags |= O_CREAT;
 544
 545                 if (is_std)
 546                         f->fd = dup(STDOUT_FILENO);
 547                 else
 548                         from_hash = file_lookup_open(f, flags);
 549         } else if (td_read(td)) {
 550                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 551                         flags |= O_RDWR;
 552                 else
 553                         flags |= O_RDONLY;
 554
 555                 if (is_std)
 556                         f->fd = dup(STDIN_FILENO);
 557                 else
 558                         from_hash = file_lookup_open(f, flags);
 559         } else { //td trim
 560                 flags |= O_RDWR;
 561                 from_hash = file_lookup_open(f, flags);
 562         }
 563
 564         if (f->fd == -1) {
 565                 char buf[FIO_VERROR_SIZE];
 566                 int __e = errno;
 567
 568                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 569                         flags &= ~FIO_O_NOATIME;
 570                         goto open_again;
 571                 }
 572                 if (__e == EMFILE && file_close_shadow_fds(td))
 573                         goto open_again;
 574
 575                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 576
 577                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 578                         log_err("fio: looks like your file system does not " \
 579                                 "support direct=1/buffered=0\n");
 580                 }
 581
 582                 td_verror(td, __e, buf);
 583         }
 584
 585         if (!from_hash && f->fd != -1) {
 586                 if (add_file_hash(f)) {
 587                         int fio_unused ret;
 588
 589                         /*
 590                          * Stash away descriptor for later close. This is to
 591                          * work-around a "feature" on Linux, where a close of
 592                          * an fd that has been opened for write will trigger
 593                          * udev to call blkid to check partitions, fs id, etc.
 594                          * That pollutes the device cache, which can slow down
 595                          * unbuffered accesses.
 596                          */
 597                         if (f->shadow_fd == -1)
 598                                 f->shadow_fd = f->fd;
 599                         else {
 600                                 /*
 601                                  * OK to ignore, we haven't done anything
 602                                  * with it
 603                                  */
 604                                 ret = generic_close_file(td, f);
 605                         }
 606                         goto open_again;
 607                 }
 608         }
 609
 610         return 0;
 611 }
 612
 613 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 614 {
 615         return get_file_size(td, f);
 616 }
 617
 618 /*
 619  * open/close all files, so that ->real_file_size gets set
 620  */
 621 static int get_file_sizes(struct thread_data *td)
 622 {
 623         struct fio_file *f;
 624         unsigned int i;
 625         int err = 0;
 626
 627         for_each_file(td, f, i) {
 628                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 629                                                                 f->file_name);
 630
 631                 if (td_io_get_file_size(td, f)) {
 632                         if (td->error != ENOENT) {
 633                                 log_err("%s\n", td->verror);
 634                                 err = 1;
 635                         }
 636                         clear_error(td);
 637                 }
 638
 639                 if (f->real_file_size == -1ULL && td->o.size)
 640                         f->real_file_size = td->o.size / td->o.nr_files;
 641         }
 642
 643         return err;
 644 }
 645
 646 struct fio_mount {
 647         struct flist_head list;
 648         const char *base;
 649         char __base[256];
 650         unsigned int key;
 651 };
 652
 653 /*
 654  * Get free number of bytes for each file on each unique mount.
 655  */
 656 static unsigned long long get_fs_free_counts(struct thread_data *td)
 657 {
 658         struct flist_head *n, *tmp;
 659         unsigned long long ret = 0;
 660         struct fio_mount *fm;
 661         FLIST_HEAD(list);
 662         struct fio_file *f;
 663         unsigned int i;
 664
 665         for_each_file(td, f, i) {
 666                 struct stat sb;
 667                 char buf[256];
 668
 669                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 670                         if (f->real_file_size != -1ULL)
 671                                 ret += f->real_file_size;
 672                         continue;
 673                 } else if (f->filetype != FIO_TYPE_FILE)
 674                         continue;
 675
 676                 strcpy(buf, f->file_name);
 677
 678                 if (stat(buf, &sb) < 0) {
 679                         if (errno != ENOENT)
 680                                 break;
 681                         strcpy(buf, ".");
 682                         if (stat(buf, &sb) < 0)
 683                                 break;
 684                 }
 685
 686                 fm = NULL;
 687                 flist_for_each(n, &list) {
 688                         fm = flist_entry(n, struct fio_mount, list);
 689                         if (fm->key == sb.st_dev)
 690                                 break;
 691
 692                         fm = NULL;
 693                 }
 694
 695                 if (fm)
 696                         continue;
 697
 698                 fm = malloc(sizeof(*fm));
 699                 strcpy(fm->__base, buf);
 700                 fm->base = basename(fm->__base);
 701                 fm->key = sb.st_dev;
 702                 flist_add(&fm->list, &list);
 703         }
 704
 705         flist_for_each_safe(n, tmp, &list) {
 706                 unsigned long long sz;
 707
 708                 fm = flist_entry(n, struct fio_mount, list);
 709                 flist_del(&fm->list);
 710
 711                 sz = get_fs_size(fm->base);
 712                 if (sz && sz != -1ULL)
 713                         ret += sz;
 714
 715                 free(fm);
 716         }
 717
 718         return ret;
 719 }
 720
 721 uint64_t get_start_offset(struct thread_data *td)
 722 {
 723         return td->o.start_offset +
 724                 (td->thread_number - 1) * td->o.offset_increment;
 725 }
 726
 727 /*
 728  * Open the files and setup files sizes, creating files if necessary.
 729  */
 730 int setup_files(struct thread_data *td)
 731 {
 732         unsigned long long total_size, extend_size;
 733         struct thread_options *o = &td->o;
 734         struct fio_file *f;
 735         unsigned int i;
 736         int err = 0, need_extend;
 737         int old_state;
 738
 739         dprint(FD_FILE, "setup files\n");
 740
 741         old_state = td->runstate;
 742         td_set_runstate(td, TD_SETTING_UP);
 743
 744         if (o->read_iolog_file)
 745                 goto done;
 746
 747         /*
 748          * if ioengine defines a setup() method, it's responsible for
 749          * opening the files and setting f->real_file_size to indicate
 750          * the valid range for that file.
 751          */
 752         if (td->io_ops->setup)
 753                 err = td->io_ops->setup(td);
 754         else
 755                 err = get_file_sizes(td);
 756
 757         if (err)
 758                 goto err_out;
 759
 760         /*
 761          * check sizes. if the files/devices do not exist and the size
 762          * isn't passed to fio, abort.
 763          */
 764         total_size = 0;
 765         for_each_file(td, f, i) {
 766                 if (f->real_file_size == -1ULL)
 767                         total_size = -1ULL;
 768                 else
 769                         total_size += f->real_file_size;
 770         }
 771
 772         if (o->fill_device)
 773                 td->fill_device_size = get_fs_free_counts(td);
 774
 775         /*
 776          * device/file sizes are zero and no size given, punt
 777          */
 778         if ((!total_size || total_size == -1ULL) && !o->size &&
 779             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 780             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 781                 log_err("%s: you need to specify size=\n", o->name);
 782                 td_verror(td, EINVAL, "total_file_size");
 783                 goto err_out;
 784         }
 785
 786         /*
 787          * now file sizes are known, so we can set ->io_size. if size= is
 788          * not given, ->io_size is just equal to ->real_file_size. if size
 789          * is given, ->io_size is size / nr_files.
 790          */
 791         extend_size = total_size = 0;
 792         need_extend = 0;
 793         for_each_file(td, f, i) {
 794                 f->file_offset = get_start_offset(td);
 795
 796                 if (!o->file_size_low) {
 797                         /*
 798                          * no file size range given, file size is equal to
 799                          * total size divided by number of files. if that is
 800                          * zero, set it to the real file size.
 801                          */
 802                         f->io_size = o->size / o->nr_files;
 803                         if (!f->io_size)
 804                                 f->io_size = f->real_file_size - f->file_offset;
 805                 } else if (f->real_file_size < o->file_size_low ||
 806                            f->real_file_size > o->file_size_high) {
 807                         if (f->file_offset > o->file_size_low)
 808                                 goto err_offset;
 809                         /*
 810                          * file size given. if it's fixed, use that. if it's a
 811                          * range, generate a random size in-between.
 812                          */
 813                         if (o->file_size_low == o->file_size_high)
 814                                 f->io_size = o->file_size_low - f->file_offset;
 815                         else {
 816                                 f->io_size = get_rand_file_size(td)
 817                                                 - f->file_offset;
 818                         }
 819                 } else
 820                         f->io_size = f->real_file_size - f->file_offset;
 821
 822                 if (f->io_size == -1ULL)
 823                         total_size = -1ULL;
 824                 else {
 825                         if (o->size_percent)
 826                                 f->io_size = (f->io_size * o->size_percent) / 100;
 827                         total_size += f->io_size;
 828                 }
 829
 830                 if (f->filetype == FIO_TYPE_FILE &&
 831                     (f->io_size + f->file_offset) > f->real_file_size &&
 832                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 833                         if (!o->create_on_open) {
 834                                 need_extend++;
 835                                 extend_size += (f->io_size + f->file_offset);
 836                         } else
 837                                 f->real_file_size = f->io_size + f->file_offset;
 838                         fio_file_set_extend(f);
 839                 }
 840         }
 841
 842         if (!o->size || o->size > total_size)
 843                 o->size = total_size;
 844
 845         /*
 846          * See if we need to extend some files
 847          */
 848         if (need_extend) {
 849                 temp_stall_ts = 1;
 850                 if (output_format == FIO_OUTPUT_NORMAL)
 851                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 852                                  " %lluMB)\n", o->name, need_extend,
 853                                         extend_size >> 20);
 854
 855                 for_each_file(td, f, i) {
 856                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 857
 858                         if (!fio_file_extend(f))
 859                                 continue;
 860
 861                         assert(f->filetype == FIO_TYPE_FILE);
 862                         fio_file_clear_extend(f);
 863                         if (!o->fill_device) {
 864                                 old_len = f->real_file_size;
 865                                 extend_len = f->io_size + f->file_offset -
 866                                                 old_len;
 867                         }
 868                         f->real_file_size = (f->io_size + f->file_offset);
 869                         err = extend_file(td, f);
 870                         if (err)
 871                                 break;
 872
 873                         err = __file_invalidate_cache(td, f, old_len,
 874                                                                 extend_len);
 875                         close(f->fd);
 876                         f->fd = -1;
 877                         if (err)
 878                                 break;
 879                 }
 880                 temp_stall_ts = 0;
 881         }
 882
 883         if (err)
 884                 goto err_out;
 885
 886         if (!o->zone_size)
 887                 o->zone_size = o->size;
 888
 889         /*
 890          * iolog already set the total io size, if we read back
 891          * stored entries.
 892          */
 893         if (!o->read_iolog_file)
 894                 td->total_io_size = o->size * o->loops;
 895
 896 done:
 897         if (o->create_only)
 898                 td->done = 1;
 899
 900         td_set_runstate(td, old_state);
 901         return 0;
 902 err_offset:
 903         log_err("%s: you need to specify valid offset=\n", o->name);
 904 err_out:
 905         td_set_runstate(td, old_state);
 906         return 1;
 907 }
 908
 909 int pre_read_files(struct thread_data *td)
 910 {
 911         struct fio_file *f;
 912         unsigned int i;
 913
 914         dprint(FD_FILE, "pre_read files\n");
 915
 916         for_each_file(td, f, i) {
 917                 pre_read_file(td, f);
 918         }
 919
 920         return 1;
 921 }
 922
 923 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 924 {
 925         unsigned int range_size, seed;
 926         unsigned long nranges;
 927         uint64_t file_size;
 928
 929         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 930         file_size = min(f->real_file_size, f->io_size);
 931
 932         nranges = (file_size + range_size - 1) / range_size;
 933
 934         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 935         if (!td->o.rand_repeatable)
 936                 seed = td->rand_seeds[4];
 937
 938         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 939                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 940         else
 941                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 942
 943         return 1;
 944 }
 945
 946 static int init_rand_distribution(struct thread_data *td)
 947 {
 948         struct fio_file *f;
 949         unsigned int i;
 950         int state;
 951
 952         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 953                 return 0;
 954
 955         state = td->runstate;
 956         td_set_runstate(td, TD_SETTING_UP);
 957         for_each_file(td, f, i)
 958                 __init_rand_distribution(td, f);
 959         td_set_runstate(td, state);
 960
 961         return 1;
 962 }
 963
 964 int init_random_map(struct thread_data *td)
 965 {
 966         unsigned long long blocks;
 967         struct fio_file *f;
 968         unsigned int i;
 969
 970         if (init_rand_distribution(td))
 971                 return 0;
 972         if (!td_random(td))
 973                 return 0;
 974
 975         for_each_file(td, f, i) {
 976                 uint64_t file_size = min(f->real_file_size, f->io_size);
 977
 978                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
 979
 980                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
 981                         unsigned long seed;
 982
 983                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
 984
 985                         if (!lfsr_init(&f->lfsr, blocks, seed, seed & 0xF))
 986                                 continue;
 987                 } else if (!td->o.norandommap) {
 988                         f->io_axmap = axmap_new(blocks);
 989                         if (f->io_axmap)
 990                                 continue;
 991                 } else if (td->o.norandommap)
 992                         continue;
 993
 994                 if (!td->o.softrandommap) {
 995                         log_err("fio: failed allocating random map. If running"
 996                                 " a large number of jobs, try the 'norandommap'"
 997                                 " option or set 'softrandommap'. Or give"
 998                                 " a larger --alloc-size to fio.\n");
 999                         return 1;
1000                 }
1001
1002                 log_info("fio: file %s failed allocating random map. Running "
1003                          "job without.\n", f->file_name);
1004         }
1005
1006         return 0;
1007 }
1008
1009 void close_files(struct thread_data *td)
1010 {
1011         struct fio_file *f;
1012         unsigned int i;
1013
1014         for_each_file(td, f, i) {
1015                 if (fio_file_open(f))
1016                         td_io_close_file(td, f);
1017         }
1018 }
1019
1020 void close_and_free_files(struct thread_data *td)
1021 {
1022         struct fio_file *f;
1023         unsigned int i;
1024
1025         dprint(FD_FILE, "close files\n");
1026
1027         for_each_file(td, f, i) {
1028                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1029                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1030                         unlink(f->file_name);
1031                 }
1032
1033                 if (fio_file_open(f))
1034                         td_io_close_file(td, f);
1035
1036                 remove_file_hash(f);
1037
1038                 sfree(f->file_name);
1039                 f->file_name = NULL;
1040                 axmap_free(f->io_axmap);
1041                 f->io_axmap = NULL;
1042                 sfree(f);
1043         }
1044
1045         td->o.filename = NULL;
1046         free(td->files);
1047         free(td->file_locks);
1048         td->files_index = 0;
1049         td->files = NULL;
1050         td->file_locks = NULL;
1051         td->o.nr_files = 0;
1052 }
1053
1054 static void get_file_type(struct fio_file *f)
1055 {
1056         struct stat sb;
1057
1058         if (!strcmp(f->file_name, "-"))
1059                 f->filetype = FIO_TYPE_PIPE;
1060         else
1061                 f->filetype = FIO_TYPE_FILE;
1062
1063         /* \\.\ is the device namespace in Windows, where every file is
1064          * a block device */
1065         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1066                 f->filetype = FIO_TYPE_BD;
1067
1068         if (!stat(f->file_name, &sb)) {
1069                 if (S_ISBLK(sb.st_mode))
1070                         f->filetype = FIO_TYPE_BD;
1071                 else if (S_ISCHR(sb.st_mode))
1072                         f->filetype = FIO_TYPE_CHAR;
1073                 else if (S_ISFIFO(sb.st_mode))
1074                         f->filetype = FIO_TYPE_PIPE;
1075         }
1076 }
1077
1078 int add_file(struct thread_data *td, const char *fname)
1079 {
1080         int cur_files = td->files_index;
1081         char file_name[PATH_MAX];
1082         struct fio_file *f;
1083         int len = 0;
1084
1085         dprint(FD_FILE, "add file %s\n", fname);
1086
1087         f = smalloc(sizeof(*f));
1088         if (!f) {
1089                 log_err("fio: smalloc OOM\n");
1090                 assert(0);
1091         }
1092
1093         f->fd = -1;
1094         f->shadow_fd = -1;
1095         fio_file_reset(td, f);
1096
1097         if (td->files_size <= td->files_index) {
1098                 unsigned int new_size = td->o.nr_files + 1;
1099
1100                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1101
1102                 td->files = realloc(td->files, new_size * sizeof(f));
1103                 if (td->files == NULL) {
1104                         log_err("fio: realloc OOM\n");
1105                         assert(0);
1106                 }
1107                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1108                         td->file_locks = realloc(td->file_locks, new_size);
1109                         if (!td->file_locks) {
1110                                 log_err("fio: realloc OOM\n");
1111                                 assert(0);
1112                         }
1113                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1114                 }
1115                 td->files_size = new_size;
1116         }
1117         td->files[cur_files] = f;
1118         f->fileno = cur_files;
1119
1120         /*
1121          * init function, io engine may not be loaded yet
1122          */
1123         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1124                 f->real_file_size = -1ULL;
1125
1126         if (td->o.directory)
1127                 len = sprintf(file_name, "%s/", td->o.directory);
1128
1129         sprintf(file_name + len, "%s", fname);
1130         f->file_name = smalloc_strdup(file_name);
1131         if (!f->file_name) {
1132                 log_err("fio: smalloc OOM\n");
1133                 assert(0);
1134         }
1135
1136         get_file_type(f);
1137
1138         switch (td->o.file_lock_mode) {
1139         case FILE_LOCK_NONE:
1140                 break;
1141         case FILE_LOCK_READWRITE:
1142                 f->rwlock = fio_rwlock_init();
1143                 break;
1144         case FILE_LOCK_EXCLUSIVE:
1145                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1146                 break;
1147         default:
1148                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1149                 assert(0);
1150         }
1151
1152         td->files_index++;
1153         if (f->filetype == FIO_TYPE_FILE)
1154                 td->nr_normal_files++;
1155
1156         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1157                                                         cur_files);
1158
1159         return cur_files;
1160 }
1161
1162 int add_file_exclusive(struct thread_data *td, const char *fname)
1163 {
1164         struct fio_file *f;
1165         unsigned int i;
1166
1167         for_each_file(td, f, i) {
1168                 if (!strcmp(f->file_name, fname))
1169                         return i;
1170         }
1171
1172         return add_file(td, fname);
1173 }
1174
1175 void get_file(struct fio_file *f)
1176 {
1177         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1178         assert(fio_file_open(f));
1179         f->references++;
1180 }
1181
1182 int put_file(struct thread_data *td, struct fio_file *f)
1183 {
1184         int f_ret = 0, ret = 0;
1185
1186         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1187
1188         if (!fio_file_open(f)) {
1189                 assert(f->fd == -1);
1190                 return 0;
1191         }
1192
1193         assert(f->references);
1194         if (--f->references)
1195                 return 0;
1196
1197         if (should_fsync(td) && td->o.fsync_on_close)
1198                 f_ret = fsync(f->fd);
1199
1200         if (td->io_ops->close_file)
1201                 ret = td->io_ops->close_file(td, f);
1202
1203         if (!ret)
1204                 ret = f_ret;
1205
1206         td->nr_open_files--;
1207         fio_file_clear_open(f);
1208         assert(f->fd == -1);
1209         return ret;
1210 }
1211
1212 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1213 {
1214         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1215                 return;
1216
1217         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1218                 if (ddir == DDIR_READ)
1219                         fio_rwlock_read(f->rwlock);
1220                 else
1221                         fio_rwlock_write(f->rwlock);
1222         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1223                 fio_mutex_down(f->lock);
1224
1225         td->file_locks[f->fileno] = td->o.file_lock_mode;
1226 }
1227
1228 void unlock_file(struct thread_data *td, struct fio_file *f)
1229 {
1230         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1231                 return;
1232
1233         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1234                 fio_rwlock_unlock(f->rwlock);
1235         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1236                 fio_mutex_up(f->lock);
1237
1238         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1239 }
1240
1241 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1242 {
1243         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1244                 unlock_file(td, f);
1245 }
1246
1247 static int recurse_dir(struct thread_data *td, const char *dirname)
1248 {
1249         struct dirent *dir;
1250         int ret = 0;
1251         DIR *D;
1252
1253         D = opendir(dirname);
1254         if (!D) {
1255                 char buf[FIO_VERROR_SIZE];
1256
1257                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1258                 td_verror(td, errno, buf);
1259                 return 1;
1260         }
1261
1262         while ((dir = readdir(D)) != NULL) {
1263                 char full_path[PATH_MAX];
1264                 struct stat sb;
1265
1266                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1267                         continue;
1268
1269                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1270
1271                 if (lstat(full_path, &sb) == -1) {
1272                         if (errno != ENOENT) {
1273                                 td_verror(td, errno, "stat");
1274                                 return 1;
1275                         }
1276                 }
1277
1278                 if (S_ISREG(sb.st_mode)) {
1279                         add_file(td, full_path);
1280                         td->o.nr_files++;
1281                         continue;
1282                 }
1283                 if (!S_ISDIR(sb.st_mode))
1284                         continue;
1285
1286                 ret = recurse_dir(td, full_path);
1287                 if (ret)
1288                         break;
1289         }
1290
1291         closedir(D);
1292         return ret;
1293 }
1294
1295 int add_dir_files(struct thread_data *td, const char *path)
1296 {
1297         int ret = recurse_dir(td, path);
1298
1299         if (!ret)
1300                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1301
1302         return ret;
1303 }
1304
1305 void dup_files(struct thread_data *td, struct thread_data *org)
1306 {
1307         struct fio_file *f;
1308         unsigned int i;
1309
1310         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1311
1312         if (!org->files)
1313                 return;
1314
1315         td->files = malloc(org->files_index * sizeof(f));
1316
1317         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1318                 td->file_locks = malloc(org->files_index);
1319
1320         for_each_file(org, f, i) {
1321                 struct fio_file *__f;
1322
1323                 __f = smalloc(sizeof(*__f));
1324                 if (!__f) {
1325                         log_err("fio: smalloc OOM\n");
1326                         assert(0);
1327                 }
1328                 __f->fd = -1;
1329                 fio_file_reset(td, __f);
1330
1331                 if (f->file_name) {
1332                         __f->file_name = smalloc_strdup(f->file_name);
1333                         if (!__f->file_name) {
1334                                 log_err("fio: smalloc OOM\n");
1335                                 assert(0);
1336                         }
1337
1338                         __f->filetype = f->filetype;
1339                 }
1340
1341                 td->files[i] = __f;
1342         }
1343 }
1344
1345 /*
1346  * Returns the index that matches the filename, or -1 if not there
1347  */
1348 int get_fileno(struct thread_data *td, const char *fname)
1349 {
1350         struct fio_file *f;
1351         unsigned int i;
1352
1353         for_each_file(td, f, i)
1354                 if (!strcmp(f->file_name, fname))
1355                         return i;
1356
1357         return -1;
1358 }
1359
1360 /*
1361  * For log usage, where we add/open/close files automatically
1362  */
1363 void free_release_files(struct thread_data *td)
1364 {
1365         close_files(td);
1366         td->files_index = 0;
1367         td->nr_normal_files = 0;
1368 }
1369
1370 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1371 {
1372         f->last_pos = f->file_offset;
1373         f->last_start = -1ULL;
1374         if (f->io_axmap)
1375                 axmap_reset(f->io_axmap);
1376         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1377                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1378 }