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 "options.h"
  15 #include "os/os.h"
  16 #include "hash.h"
  17 #include "lib/axmap.h"
  18
  19 #ifdef CONFIG_LINUX_FALLOCATE
  20 #include <linux/falloc.h>
  21 #endif
  22
  23 static int root_warn;
  24
  25 static FLIST_HEAD(filename_list);
  26
  27 static inline void clear_error(struct thread_data *td)
  28 {
  29         td->error = 0;
  30         td->verror[0] = '\0';
  31 }
  32
  33 /*
  34  * Leaves f->fd open on success, caller must close
  35  */
  36 static int extend_file(struct thread_data *td, struct fio_file *f)
  37 {
  38         int r, new_layout = 0, unlink_file = 0, flags;
  39         unsigned long long left;
  40         unsigned int bs;
  41         char *b;
  42
  43         if (read_only) {
  44                 log_err("fio: refusing extend of file due to read-only\n");
  45                 return 0;
  46         }
  47
  48         /*
  49          * check if we need to lay the file out complete again. fio
  50          * does that for operations involving reads, or for writes
  51          * where overwrite is set
  52          */
  53         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
  54             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
  55                 new_layout = 1;
  56         if (td_write(td) && !td->o.overwrite)
  57                 unlink_file = 1;
  58
  59         if (unlink_file || new_layout) {
  60                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  61                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
  62                         td_verror(td, errno, "unlink");
  63                         return 1;
  64                 }
  65         }
  66
  67         flags = O_WRONLY | O_CREAT;
  68         if (new_layout)
  69                 flags |= O_TRUNC;
  70
  71         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  72         f->fd = open(f->file_name, flags, 0644);
  73         if (f->fd < 0) {
  74                 td_verror(td, errno, "open");
  75                 return 1;
  76         }
  77
  78 #ifdef CONFIG_POSIX_FALLOCATE
  79         if (!td->o.fill_device) {
  80                 switch (td->o.fallocate_mode) {
  81                 case FIO_FALLOCATE_NONE:
  82                         break;
  83                 case FIO_FALLOCATE_POSIX:
  84                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  85                                  f->file_name,
  86                                  (unsigned long long) f->real_file_size);
  87
  88                         r = posix_fallocate(f->fd, 0, f->real_file_size);
  89                         if (r > 0) {
  90                                 log_err("fio: posix_fallocate fails: %s\n",
  91                                                 strerror(r));
  92                         }
  93                         break;
  94 #ifdef CONFIG_LINUX_FALLOCATE
  95                 case FIO_FALLOCATE_KEEP_SIZE:
  96                         dprint(FD_FILE,
  97                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
  98                                 "file %s size %llu\n", f->file_name,
  99                                 (unsigned long long) f->real_file_size);
 100
 101                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
 102                                         f->real_file_size);
 103                         if (r != 0)
 104                                 td_verror(td, errno, "fallocate");
 105
 106                         break;
 107 #endif /* CONFIG_LINUX_FALLOCATE */
 108                 default:
 109                         log_err("fio: unknown fallocate mode: %d\n",
 110                                 td->o.fallocate_mode);
 111                         assert(0);
 112                 }
 113         }
 114 #endif /* CONFIG_POSIX_FALLOCATE */
 115
 116         if (!new_layout)
 117                 goto done;
 118
 119         /*
 120          * The size will be -1ULL when fill_device is used, so don't truncate
 121          * or fallocate this file, just write it
 122          */
 123         if (!td->o.fill_device) {
 124                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 125                                         (unsigned long long) f->real_file_size);
 126                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 127                         if (errno != EFBIG) {
 128                                 td_verror(td, errno, "ftruncate");
 129                                 goto err;
 130                         }
 131                 }
 132         }
 133
 134         b = malloc(td->o.max_bs[DDIR_WRITE]);
 135
 136         left = f->real_file_size;
 137         while (left && !td->terminate) {
 138                 bs = td->o.max_bs[DDIR_WRITE];
 139                 if (bs > left)
 140                         bs = left;
 141
 142                 fill_io_buffer(td, b, bs, bs);
 143
 144                 r = write(f->fd, b, bs);
 145
 146                 if (r > 0) {
 147                         left -= r;
 148                         continue;
 149                 } else {
 150                         if (r < 0) {
 151                                 int __e = errno;
 152
 153                                 if (__e == ENOSPC) {
 154                                         if (td->o.fill_device)
 155                                                 break;
 156                                         log_info("fio: ENOSPC on laying out "
 157                                                  "file, stopping\n");
 158                                         break;
 159                                 }
 160                                 td_verror(td, errno, "write");
 161                         } else
 162                                 td_verror(td, EIO, "write");
 163
 164                         break;
 165                 }
 166         }
 167
 168         if (td->terminate) {
 169                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 170                 unlink(f->file_name);
 171         } else if (td->o.create_fsync) {
 172                 if (fsync(f->fd) < 0) {
 173                         td_verror(td, errno, "fsync");
 174                         goto err;
 175                 }
 176         }
 177         if (td->o.fill_device && !td_write(td)) {
 178                 fio_file_clear_size_known(f);
 179                 if (td_io_get_file_size(td, f))
 180                         goto err;
 181                 if (f->io_size > f->real_file_size)
 182                         f->io_size = f->real_file_size;
 183         }
 184
 185         free(b);
 186 done:
 187         return 0;
 188 err:
 189         close(f->fd);
 190         f->fd = -1;
 191         return 1;
 192 }
 193
 194 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 195 {
 196         int r, did_open = 0, old_runstate;
 197         unsigned long long left;
 198         unsigned int bs;
 199         char *b;
 200
 201         if (td->io_ops->flags & FIO_PIPEIO)
 202                 return 0;
 203
 204         if (!fio_file_open(f)) {
 205                 if (td->io_ops->open_file(td, f)) {
 206                         log_err("fio: cannot pre-read, failed to open file\n");
 207                         return 1;
 208                 }
 209                 did_open = 1;
 210         }
 211
 212         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 213
 214         bs = td->o.max_bs[DDIR_READ];
 215         b = malloc(bs);
 216         memset(b, 0, bs);
 217
 218         lseek(f->fd, f->file_offset, SEEK_SET);
 219         left = f->io_size;
 220
 221         while (left && !td->terminate) {
 222                 if (bs > left)
 223                         bs = left;
 224
 225                 r = read(f->fd, b, bs);
 226
 227                 if (r == (int) bs) {
 228                         left -= bs;
 229                         continue;
 230                 } else {
 231                         td_verror(td, EIO, "pre_read");
 232                         break;
 233                 }
 234         }
 235
 236         td_restore_runstate(td, old_runstate);
 237
 238         if (did_open)
 239                 td->io_ops->close_file(td, f);
 240         free(b);
 241         return 0;
 242 }
 243
 244 static unsigned long long get_rand_file_size(struct thread_data *td)
 245 {
 246         unsigned long long ret, sized;
 247         unsigned long r;
 248
 249         if (td->o.use_os_rand) {
 250                 r = os_random_long(&td->file_size_state);
 251                 sized = td->o.file_size_high - td->o.file_size_low;
 252                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 253         } else {
 254                 r = __rand(&td->__file_size_state);
 255                 sized = td->o.file_size_high - td->o.file_size_low;
 256                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 257         }
 258
 259         ret += td->o.file_size_low;
 260         ret -= (ret % td->o.rw_min_bs);
 261         return ret;
 262 }
 263
 264 static int file_size(struct thread_data *td, struct fio_file *f)
 265 {
 266         struct stat st;
 267
 268         if (stat(f->file_name, &st) == -1) {
 269                 td_verror(td, errno, "fstat");
 270                 return 1;
 271         }
 272
 273         f->real_file_size = st.st_size;
 274         return 0;
 275 }
 276
 277 static int bdev_size(struct thread_data *td, struct fio_file *f)
 278 {
 279         unsigned long long bytes = 0;
 280         int r;
 281
 282         if (td->io_ops->open_file(td, f)) {
 283                 log_err("fio: failed opening blockdev %s for size check\n",
 284                         f->file_name);
 285                 return 1;
 286         }
 287
 288         r = blockdev_size(f, &bytes);
 289         if (r) {
 290                 td_verror(td, r, "blockdev_size");
 291                 goto err;
 292         }
 293
 294         if (!bytes) {
 295                 log_err("%s: zero sized block device?\n", f->file_name);
 296                 goto err;
 297         }
 298
 299         f->real_file_size = bytes;
 300         td->io_ops->close_file(td, f);
 301         return 0;
 302 err:
 303         td->io_ops->close_file(td, f);
 304         return 1;
 305 }
 306
 307 static int char_size(struct thread_data *td, struct fio_file *f)
 308 {
 309 #ifdef FIO_HAVE_CHARDEV_SIZE
 310         unsigned long long bytes = 0;
 311         int r;
 312
 313         if (td->io_ops->open_file(td, f)) {
 314                 log_err("fio: failed opening blockdev %s for size check\n",
 315                         f->file_name);
 316                 return 1;
 317         }
 318
 319         r = chardev_size(f, &bytes);
 320         if (r) {
 321                 td_verror(td, r, "chardev_size");
 322                 goto err;
 323         }
 324
 325         if (!bytes) {
 326                 log_err("%s: zero sized char device?\n", f->file_name);
 327                 goto err;
 328         }
 329
 330         f->real_file_size = bytes;
 331         td->io_ops->close_file(td, f);
 332         return 0;
 333 err:
 334         td->io_ops->close_file(td, f);
 335         return 1;
 336 #else
 337         f->real_file_size = -1ULL;
 338         return 0;
 339 #endif
 340 }
 341
 342 static int get_file_size(struct thread_data *td, struct fio_file *f)
 343 {
 344         int ret = 0;
 345
 346         if (fio_file_size_known(f))
 347                 return 0;
 348
 349         if (f->filetype == FIO_TYPE_FILE)
 350                 ret = file_size(td, f);
 351         else if (f->filetype == FIO_TYPE_BD)
 352                 ret = bdev_size(td, f);
 353         else if (f->filetype == FIO_TYPE_CHAR)
 354                 ret = char_size(td, f);
 355         else
 356                 f->real_file_size = -1;
 357
 358         if (ret)
 359                 return ret;
 360
 361         if (f->file_offset > f->real_file_size) {
 362                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 363                                         (unsigned long long) f->file_offset,
 364                                         (unsigned long long) f->real_file_size);
 365                 return 1;
 366         }
 367
 368         fio_file_set_size_known(f);
 369         return 0;
 370 }
 371
 372 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 373                                    unsigned long long off,
 374                                    unsigned long long len)
 375 {
 376         int ret = 0;
 377
 378         if (len == -1ULL)
 379                 len = f->io_size;
 380         if (off == -1ULL)
 381                 off = f->file_offset;
 382
 383         if (len == -1ULL || off == -1ULL)
 384                 return 0;
 385
 386         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 387                                                                 len);
 388
 389         /*
 390          * FIXME: add blockdev flushing too
 391          */
 392         if (f->mmap_ptr) {
 393                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 394 #ifdef FIO_MADV_FREE
 395                 if (f->filetype == FIO_TYPE_BD)
 396                         (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 397 #endif
 398         } else if (f->filetype == FIO_TYPE_FILE) {
 399                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 400         } else if (f->filetype == FIO_TYPE_BD) {
 401                 ret = blockdev_invalidate_cache(f);
 402                 if (ret < 0 && errno == EACCES && geteuid()) {
 403                         if (!root_warn) {
 404                                 log_err("fio: only root may flush block "
 405                                         "devices. Cache flush bypassed!\n");
 406                                 root_warn = 1;
 407                         }
 408                         ret = 0;
 409                 }
 410         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 411                 ret = 0;
 412
 413         if (ret < 0) {
 414                 td_verror(td, errno, "invalidate_cache");
 415                 return 1;
 416         } else if (ret > 0) {
 417                 td_verror(td, ret, "invalidate_cache");
 418                 return 1;
 419         }
 420
 421         return ret;
 422
 423 }
 424
 425 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 426 {
 427         if (!fio_file_open(f))
 428                 return 0;
 429
 430         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 431 }
 432
 433 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 434 {
 435         int ret = 0;
 436
 437         dprint(FD_FILE, "fd close %s\n", f->file_name);
 438
 439         remove_file_hash(f);
 440
 441         if (close(f->fd) < 0)
 442                 ret = errno;
 443
 444         f->fd = -1;
 445
 446         if (f->shadow_fd != -1) {
 447                 close(f->shadow_fd);
 448                 f->shadow_fd = -1;
 449         }
 450
 451         f->engine_data = 0;
 452         return ret;
 453 }
 454
 455 int file_lookup_open(struct fio_file *f, int flags)
 456 {
 457         struct fio_file *__f;
 458         int from_hash;
 459
 460         __f = lookup_file_hash(f->file_name);
 461         if (__f) {
 462                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 463                 /*
 464                  * racy, need the __f->lock locked
 465                  */
 466                 f->lock = __f->lock;
 467                 from_hash = 1;
 468         } else {
 469                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 470                 from_hash = 0;
 471         }
 472
 473         f->fd = open(f->file_name, flags, 0600);
 474         return from_hash;
 475 }
 476
 477 static int file_close_shadow_fds(struct thread_data *td)
 478 {
 479         struct fio_file *f;
 480         int num_closed = 0;
 481         unsigned int i;
 482
 483         for_each_file(td, f, i) {
 484                 if (f->shadow_fd == -1)
 485                         continue;
 486
 487                 close(f->shadow_fd);
 488                 f->shadow_fd = -1;
 489                 num_closed++;
 490         }
 491
 492         return num_closed;
 493 }
 494
 495 int generic_open_file(struct thread_data *td, struct fio_file *f)
 496 {
 497         int is_std = 0;
 498         int flags = 0;
 499         int from_hash = 0;
 500
 501         dprint(FD_FILE, "fd open %s\n", f->file_name);
 502
 503         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 504                 log_err("fio: trim only applies to block device\n");
 505                 return 1;
 506         }
 507
 508         if (!strcmp(f->file_name, "-")) {
 509                 if (td_rw(td)) {
 510                         log_err("fio: can't read/write to stdin/out\n");
 511                         return 1;
 512                 }
 513                 is_std = 1;
 514
 515                 /*
 516                  * move output logging to stderr, if we are writing to stdout
 517                  */
 518                 if (td_write(td))
 519                         f_out = stderr;
 520         }
 521
 522         if (td_trim(td))
 523                 goto skip_flags;
 524         if (td->o.odirect)
 525                 flags |= OS_O_DIRECT;
 526         if (td->o.oatomic) {
 527                 if (!FIO_O_ATOMIC) {
 528                         td_verror(td, EINVAL, "OS does not support atomic IO");
 529                         return 1;
 530                 }
 531                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 532         }
 533         if (td->o.sync_io)
 534                 flags |= O_SYNC;
 535         if (td->o.create_on_open)
 536                 flags |= O_CREAT;
 537 skip_flags:
 538         if (f->filetype != FIO_TYPE_FILE)
 539                 flags |= FIO_O_NOATIME;
 540
 541 open_again:
 542         if (td_write(td)) {
 543                 if (!read_only)
 544                         flags |= O_RDWR;
 545
 546                 if (f->filetype == FIO_TYPE_FILE)
 547                         flags |= O_CREAT;
 548
 549                 if (is_std)
 550                         f->fd = dup(STDOUT_FILENO);
 551                 else
 552                         from_hash = file_lookup_open(f, flags);
 553         } else if (td_read(td)) {
 554                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 555                         flags |= O_RDWR;
 556                 else
 557                         flags |= O_RDONLY;
 558
 559                 if (is_std)
 560                         f->fd = dup(STDIN_FILENO);
 561                 else
 562                         from_hash = file_lookup_open(f, flags);
 563         } else { //td trim
 564                 flags |= O_RDWR;
 565                 from_hash = file_lookup_open(f, flags);
 566         }
 567
 568         if (f->fd == -1) {
 569                 char buf[FIO_VERROR_SIZE];
 570                 int __e = errno;
 571
 572                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 573                         flags &= ~FIO_O_NOATIME;
 574                         goto open_again;
 575                 }
 576                 if (__e == EMFILE && file_close_shadow_fds(td))
 577                         goto open_again;
 578
 579                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 580
 581                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 582                         log_err("fio: looks like your file system does not " \
 583                                 "support direct=1/buffered=0\n");
 584                 }
 585
 586                 td_verror(td, __e, buf);
 587         }
 588
 589         if (!from_hash && f->fd != -1) {
 590                 if (add_file_hash(f)) {
 591                         int fio_unused ret;
 592
 593                         /*
 594                          * Stash away descriptor for later close. This is to
 595                          * work-around a "feature" on Linux, where a close of
 596                          * an fd that has been opened for write will trigger
 597                          * udev to call blkid to check partitions, fs id, etc.
 598                          * That pollutes the device cache, which can slow down
 599                          * unbuffered accesses.
 600                          */
 601                         if (f->shadow_fd == -1)
 602                                 f->shadow_fd = f->fd;
 603                         else {
 604                                 /*
 605                                  * OK to ignore, we haven't done anything
 606                                  * with it
 607                                  */
 608                                 ret = generic_close_file(td, f);
 609                         }
 610                         goto open_again;
 611                 }
 612         }
 613
 614         return 0;
 615 }
 616
 617 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 618 {
 619         return get_file_size(td, f);
 620 }
 621
 622 /*
 623  * open/close all files, so that ->real_file_size gets set
 624  */
 625 static int get_file_sizes(struct thread_data *td)
 626 {
 627         struct fio_file *f;
 628         unsigned int i;
 629         int err = 0;
 630
 631         for_each_file(td, f, i) {
 632                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 633                                                                 f->file_name);
 634
 635                 if (td_io_get_file_size(td, f)) {
 636                         if (td->error != ENOENT) {
 637                                 log_err("%s\n", td->verror);
 638                                 err = 1;
 639                         }
 640                         clear_error(td);
 641                 }
 642
 643                 if (f->real_file_size == -1ULL && td->o.size)
 644                         f->real_file_size = td->o.size / td->o.nr_files;
 645         }
 646
 647         return err;
 648 }
 649
 650 struct fio_mount {
 651         struct flist_head list;
 652         const char *base;
 653         char __base[256];
 654         unsigned int key;
 655 };
 656
 657 /*
 658  * Get free number of bytes for each file on each unique mount.
 659  */
 660 static unsigned long long get_fs_free_counts(struct thread_data *td)
 661 {
 662         struct flist_head *n, *tmp;
 663         unsigned long long ret = 0;
 664         struct fio_mount *fm;
 665         FLIST_HEAD(list);
 666         struct fio_file *f;
 667         unsigned int i;
 668
 669         for_each_file(td, f, i) {
 670                 struct stat sb;
 671                 char buf[256];
 672
 673                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 674                         if (f->real_file_size != -1ULL)
 675                                 ret += f->real_file_size;
 676                         continue;
 677                 } else if (f->filetype != FIO_TYPE_FILE)
 678                         continue;
 679
 680                 strcpy(buf, f->file_name);
 681
 682                 if (stat(buf, &sb) < 0) {
 683                         if (errno != ENOENT)
 684                                 break;
 685                         strcpy(buf, ".");
 686                         if (stat(buf, &sb) < 0)
 687                                 break;
 688                 }
 689
 690                 fm = NULL;
 691                 flist_for_each(n, &list) {
 692                         fm = flist_entry(n, struct fio_mount, list);
 693                         if (fm->key == sb.st_dev)
 694                                 break;
 695
 696                         fm = NULL;
 697                 }
 698
 699                 if (fm)
 700                         continue;
 701
 702                 fm = malloc(sizeof(*fm));
 703                 strcpy(fm->__base, buf);
 704                 fm->base = basename(fm->__base);
 705                 fm->key = sb.st_dev;
 706                 flist_add(&fm->list, &list);
 707         }
 708
 709         flist_for_each_safe(n, tmp, &list) {
 710                 unsigned long long sz;
 711
 712                 fm = flist_entry(n, struct fio_mount, list);
 713                 flist_del(&fm->list);
 714
 715                 sz = get_fs_size(fm->base);
 716                 if (sz && sz != -1ULL)
 717                         ret += sz;
 718
 719                 free(fm);
 720         }
 721
 722         return ret;
 723 }
 724
 725 uint64_t get_start_offset(struct thread_data *td)
 726 {
 727         return td->o.start_offset +
 728                 (td->thread_number - 1) * td->o.offset_increment;
 729 }
 730
 731 /*
 732  * Open the files and setup files sizes, creating files if necessary.
 733  */
 734 int setup_files(struct thread_data *td)
 735 {
 736         unsigned long long total_size, extend_size;
 737         struct thread_options *o = &td->o;
 738         struct fio_file *f;
 739         unsigned int i, nr_fs_extra = 0;
 740         int err = 0, need_extend;
 741         int old_state;
 742         const unsigned int bs = td_min_bs(td);
 743         uint64_t fs = 0;
 744
 745         dprint(FD_FILE, "setup files\n");
 746
 747         old_state = td_bump_runstate(td, TD_SETTING_UP);
 748
 749         if (o->read_iolog_file)
 750                 goto done;
 751
 752         /*
 753          * if ioengine defines a setup() method, it's responsible for
 754          * opening the files and setting f->real_file_size to indicate
 755          * the valid range for that file.
 756          */
 757         if (td->io_ops->setup)
 758                 err = td->io_ops->setup(td);
 759         else
 760                 err = get_file_sizes(td);
 761
 762         if (err)
 763                 goto err_out;
 764
 765         /*
 766          * check sizes. if the files/devices do not exist and the size
 767          * isn't passed to fio, abort.
 768          */
 769         total_size = 0;
 770         for_each_file(td, f, i) {
 771                 if (f->real_file_size == -1ULL)
 772                         total_size = -1ULL;
 773                 else
 774                         total_size += f->real_file_size;
 775         }
 776
 777         if (o->fill_device)
 778                 td->fill_device_size = get_fs_free_counts(td);
 779
 780         /*
 781          * device/file sizes are zero and no size given, punt
 782          */
 783         if ((!total_size || total_size == -1ULL) && !o->size &&
 784             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 785             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 786                 log_err("%s: you need to specify size=\n", o->name);
 787                 td_verror(td, EINVAL, "total_file_size");
 788                 goto err_out;
 789         }
 790
 791         /*
 792          * Calculate per-file size and potential extra size for the
 793          * first files, if needed.
 794          */
 795         if (!o->file_size_low) {
 796                 uint64_t all_fs;
 797
 798                 fs = o->size / o->nr_files;
 799                 all_fs = fs * o->nr_files;
 800
 801                 if (all_fs < o->size)
 802                         nr_fs_extra = (o->size - all_fs) / bs;
 803         }
 804
 805         /*
 806          * now file sizes are known, so we can set ->io_size. if size= is
 807          * not given, ->io_size is just equal to ->real_file_size. if size
 808          * is given, ->io_size is size / nr_files.
 809          */
 810         extend_size = total_size = 0;
 811         need_extend = 0;
 812         for_each_file(td, f, i) {
 813                 f->file_offset = get_start_offset(td);
 814
 815                 if (!o->file_size_low) {
 816                         /*
 817                          * no file size range given, file size is equal to
 818                          * total size divided by number of files. If that is
 819                          * zero, set it to the real file size. If the size
 820                          * doesn't divide nicely with the min blocksize,
 821                          * make the first files bigger.
 822                          */
 823                         f->io_size = fs;
 824                         if (nr_fs_extra) {
 825                                 nr_fs_extra--;
 826                                 f->io_size += bs;
 827                         }
 828
 829                         if (!f->io_size)
 830                                 f->io_size = f->real_file_size - f->file_offset;
 831                 } else if (f->real_file_size < o->file_size_low ||
 832                            f->real_file_size > o->file_size_high) {
 833                         if (f->file_offset > o->file_size_low)
 834                                 goto err_offset;
 835                         /*
 836                          * file size given. if it's fixed, use that. if it's a
 837                          * range, generate a random size in-between.
 838                          */
 839                         if (o->file_size_low == o->file_size_high)
 840                                 f->io_size = o->file_size_low - f->file_offset;
 841                         else {
 842                                 f->io_size = get_rand_file_size(td)
 843                                                 - f->file_offset;
 844                         }
 845                 } else
 846                         f->io_size = f->real_file_size - f->file_offset;
 847
 848                 if (f->io_size == -1ULL)
 849                         total_size = -1ULL;
 850                 else {
 851                         if (o->size_percent)
 852                                 f->io_size = (f->io_size * o->size_percent) / 100;
 853                         total_size += f->io_size;
 854                 }
 855
 856                 if (f->filetype == FIO_TYPE_FILE &&
 857                     (f->io_size + f->file_offset) > f->real_file_size &&
 858                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 859                         if (!o->create_on_open) {
 860                                 need_extend++;
 861                                 extend_size += (f->io_size + f->file_offset);
 862                         } else
 863                                 f->real_file_size = f->io_size + f->file_offset;
 864                         fio_file_set_extend(f);
 865                 }
 866         }
 867
 868         if (!o->size || o->size > total_size)
 869                 o->size = total_size;
 870
 871         /*
 872          * See if we need to extend some files
 873          */
 874         if (need_extend) {
 875                 temp_stall_ts = 1;
 876                 if (output_format == FIO_OUTPUT_NORMAL)
 877                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 878                                  " %lluMB)\n", o->name, need_extend,
 879                                         extend_size >> 20);
 880
 881                 for_each_file(td, f, i) {
 882                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 883
 884                         if (!fio_file_extend(f))
 885                                 continue;
 886
 887                         assert(f->filetype == FIO_TYPE_FILE);
 888                         fio_file_clear_extend(f);
 889                         if (!o->fill_device) {
 890                                 old_len = f->real_file_size;
 891                                 extend_len = f->io_size + f->file_offset -
 892                                                 old_len;
 893                         }
 894                         f->real_file_size = (f->io_size + f->file_offset);
 895                         err = extend_file(td, f);
 896                         if (err)
 897                                 break;
 898
 899                         err = __file_invalidate_cache(td, f, old_len,
 900                                                                 extend_len);
 901                         close(f->fd);
 902                         f->fd = -1;
 903                         if (err)
 904                                 break;
 905                 }
 906                 temp_stall_ts = 0;
 907         }
 908
 909         if (err)
 910                 goto err_out;
 911
 912         if (!o->zone_size)
 913                 o->zone_size = o->size;
 914
 915         /*
 916          * iolog already set the total io size, if we read back
 917          * stored entries.
 918          */
 919         if (!o->read_iolog_file)
 920                 td->total_io_size = o->size * o->loops;
 921
 922 done:
 923         if (o->create_only)
 924                 td->done = 1;
 925
 926         td_restore_runstate(td, old_state);
 927         return 0;
 928 err_offset:
 929         log_err("%s: you need to specify valid offset=\n", o->name);
 930 err_out:
 931         td_restore_runstate(td, old_state);
 932         return 1;
 933 }
 934
 935 int pre_read_files(struct thread_data *td)
 936 {
 937         struct fio_file *f;
 938         unsigned int i;
 939
 940         dprint(FD_FILE, "pre_read files\n");
 941
 942         for_each_file(td, f, i) {
 943                 pre_read_file(td, f);
 944         }
 945
 946         return 1;
 947 }
 948
 949 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 950 {
 951         unsigned int range_size, seed;
 952         unsigned long nranges;
 953         uint64_t file_size;
 954
 955         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 956         file_size = min(f->real_file_size, f->io_size);
 957
 958         nranges = (file_size + range_size - 1) / range_size;
 959
 960         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 961         if (!td->o.rand_repeatable)
 962                 seed = td->rand_seeds[4];
 963
 964         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 965                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 966         else
 967                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 968
 969         return 1;
 970 }
 971
 972 static int init_rand_distribution(struct thread_data *td)
 973 {
 974         struct fio_file *f;
 975         unsigned int i;
 976         int state;
 977
 978         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 979                 return 0;
 980
 981         state = td_bump_runstate(td, TD_SETTING_UP);
 982
 983         for_each_file(td, f, i)
 984                 __init_rand_distribution(td, f);
 985
 986         td_restore_runstate(td, state);
 987
 988         return 1;
 989 }
 990
 991 int init_random_map(struct thread_data *td)
 992 {
 993         unsigned long long blocks;
 994         struct fio_file *f;
 995         unsigned int i;
 996
 997         if (init_rand_distribution(td))
 998                 return 0;
 999         if (!td_random(td))
1000                 return 0;
1001
1002         for_each_file(td, f, i) {
1003                 uint64_t file_size = min(f->real_file_size, f->io_size);
1004
1005                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1006
1007                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1008                         unsigned long seed;
1009
1010                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1011
1012                         if (!lfsr_init(&f->lfsr, blocks, seed, 0))
1013                                 continue;
1014                 } else if (!td->o.norandommap) {
1015                         f->io_axmap = axmap_new(blocks);
1016                         if (f->io_axmap)
1017                                 continue;
1018                 } else if (td->o.norandommap)
1019                         continue;
1020
1021                 if (!td->o.softrandommap) {
1022                         log_err("fio: failed allocating random map. If running"
1023                                 " a large number of jobs, try the 'norandommap'"
1024                                 " option or set 'softrandommap'. Or give"
1025                                 " a larger --alloc-size to fio.\n");
1026                         return 1;
1027                 }
1028
1029                 log_info("fio: file %s failed allocating random map. Running "
1030                          "job without.\n", f->file_name);
1031         }
1032
1033         return 0;
1034 }
1035
1036 void close_files(struct thread_data *td)
1037 {
1038         struct fio_file *f;
1039         unsigned int i;
1040
1041         for_each_file(td, f, i) {
1042                 if (fio_file_open(f))
1043                         td_io_close_file(td, f);
1044         }
1045 }
1046
1047 void close_and_free_files(struct thread_data *td)
1048 {
1049         struct fio_file *f;
1050         unsigned int i;
1051
1052         dprint(FD_FILE, "close files\n");
1053
1054         for_each_file(td, f, i) {
1055                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1056                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1057                         unlink(f->file_name);
1058                 }
1059
1060                 if (fio_file_open(f))
1061                         td_io_close_file(td, f);
1062
1063                 remove_file_hash(f);
1064
1065                 sfree(f->file_name);
1066                 f->file_name = NULL;
1067                 axmap_free(f->io_axmap);
1068                 f->io_axmap = NULL;
1069                 sfree(f);
1070         }
1071
1072         td->o.filename = NULL;
1073         free(td->files);
1074         free(td->file_locks);
1075         td->files_index = 0;
1076         td->files = NULL;
1077         td->file_locks = NULL;
1078         td->o.file_lock_mode = FILE_LOCK_NONE;
1079         td->o.nr_files = 0;
1080 }
1081
1082 static void get_file_type(struct fio_file *f)
1083 {
1084         struct stat sb;
1085
1086         if (!strcmp(f->file_name, "-"))
1087                 f->filetype = FIO_TYPE_PIPE;
1088         else
1089                 f->filetype = FIO_TYPE_FILE;
1090
1091         /* \\.\ is the device namespace in Windows, where every file is
1092          * a block device */
1093         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1094                 f->filetype = FIO_TYPE_BD;
1095
1096         if (!stat(f->file_name, &sb)) {
1097                 if (S_ISBLK(sb.st_mode))
1098                         f->filetype = FIO_TYPE_BD;
1099                 else if (S_ISCHR(sb.st_mode))
1100                         f->filetype = FIO_TYPE_CHAR;
1101                 else if (S_ISFIFO(sb.st_mode))
1102                         f->filetype = FIO_TYPE_PIPE;
1103         }
1104 }
1105
1106 static void set_already_allocated(const char *fname) {
1107         struct file_name *fn;
1108
1109         fn = malloc(sizeof(struct file_name));
1110         fn->filename = strdup(fname);
1111         flist_add_tail(&fn->list, &filename_list);
1112 }
1113
1114 static int is_already_allocated(const char *fname)
1115 {
1116         struct flist_head *entry;
1117         char *filename;
1118
1119         if (!flist_empty(&filename_list))
1120         {
1121                 flist_for_each(entry, &filename_list) {
1122                         filename = flist_entry(entry, struct file_name, list)->filename;
1123
1124                         if (strcmp(filename, fname) == 0)
1125                                 return 1;
1126                 }
1127         }
1128
1129         return 0;
1130 }
1131
1132 static void free_already_allocated() {
1133         struct flist_head *entry, *tmp;
1134         struct file_name *fn;
1135
1136         if (!flist_empty(&filename_list))
1137         {
1138                 flist_for_each_safe(entry, tmp, &filename_list) {
1139                         fn = flist_entry(entry, struct file_name, list);
1140                         free(fn->filename);
1141                         flist_del(&fn->list);
1142                         free(fn);
1143                 }
1144         }
1145 }
1146
1147 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1148 {
1149         int cur_files = td->files_index;
1150         char file_name[PATH_MAX];
1151         struct fio_file *f;
1152         int len = 0;
1153
1154         dprint(FD_FILE, "add file %s\n", fname);
1155
1156         if (td->o.directory)
1157                 len = set_name_idx(file_name, td->o.directory, numjob);
1158
1159         sprintf(file_name + len, "%s", fname);
1160
1161         /* clean cloned siblings using existing files */
1162         if (numjob && is_already_allocated(file_name))
1163                 return 0;
1164
1165         f = smalloc(sizeof(*f));
1166         if (!f) {
1167                 log_err("fio: smalloc OOM\n");
1168                 assert(0);
1169         }
1170
1171         f->fd = -1;
1172         f->shadow_fd = -1;
1173         fio_file_reset(td, f);
1174
1175         if (td->files_size <= td->files_index) {
1176                 unsigned int new_size = td->o.nr_files + 1;
1177
1178                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1179
1180                 td->files = realloc(td->files, new_size * sizeof(f));
1181                 if (td->files == NULL) {
1182                         log_err("fio: realloc OOM\n");
1183                         assert(0);
1184                 }
1185                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1186                         td->file_locks = realloc(td->file_locks, new_size);
1187                         if (!td->file_locks) {
1188                                 log_err("fio: realloc OOM\n");
1189                                 assert(0);
1190                         }
1191                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1192                 }
1193                 td->files_size = new_size;
1194         }
1195         td->files[cur_files] = f;
1196         f->fileno = cur_files;
1197
1198         /*
1199          * init function, io engine may not be loaded yet
1200          */
1201         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1202                 f->real_file_size = -1ULL;
1203
1204         f->file_name = smalloc_strdup(file_name);
1205         if (!f->file_name) {
1206                 log_err("fio: smalloc OOM\n");
1207                 assert(0);
1208         }
1209
1210         get_file_type(f);
1211
1212         switch (td->o.file_lock_mode) {
1213         case FILE_LOCK_NONE:
1214                 break;
1215         case FILE_LOCK_READWRITE:
1216                 f->rwlock = fio_rwlock_init();
1217                 break;
1218         case FILE_LOCK_EXCLUSIVE:
1219                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1220                 break;
1221         default:
1222                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1223                 assert(0);
1224         }
1225
1226         td->files_index++;
1227         if (f->filetype == FIO_TYPE_FILE)
1228                 td->nr_normal_files++;
1229
1230         set_already_allocated(file_name);
1231
1232         /*
1233          * For adding files after the fact - if openfiles= isn't
1234          * given as an option, ensure we allow at least one file open
1235          */
1236         if (!td->o.open_files)
1237                 td->o.open_files = 1;
1238
1239         if (inc)
1240                 td->o.nr_files++;
1241
1242         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1243                                                         cur_files);
1244
1245         return cur_files;
1246 }
1247
1248 int add_file_exclusive(struct thread_data *td, const char *fname)
1249 {
1250         struct fio_file *f;
1251         unsigned int i;
1252
1253         for_each_file(td, f, i) {
1254                 if (!strcmp(f->file_name, fname))
1255                         return i;
1256         }
1257
1258         return add_file(td, fname, 0, 1);
1259 }
1260
1261 void get_file(struct fio_file *f)
1262 {
1263         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1264         assert(fio_file_open(f));
1265         f->references++;
1266 }
1267
1268 int put_file(struct thread_data *td, struct fio_file *f)
1269 {
1270         int f_ret = 0, ret = 0;
1271
1272         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1273
1274         if (!fio_file_open(f)) {
1275                 assert(f->fd == -1);
1276                 return 0;
1277         }
1278
1279         assert(f->references);
1280         if (--f->references)
1281                 return 0;
1282
1283         if (should_fsync(td) && td->o.fsync_on_close)
1284                 f_ret = fsync(f->fd);
1285
1286         if (td->io_ops->close_file)
1287                 ret = td->io_ops->close_file(td, f);
1288
1289         if (!ret)
1290                 ret = f_ret;
1291
1292         td->nr_open_files--;
1293         fio_file_clear_open(f);
1294         assert(f->fd == -1);
1295         return ret;
1296 }
1297
1298 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1299 {
1300         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1301                 return;
1302
1303         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1304                 if (ddir == DDIR_READ)
1305                         fio_rwlock_read(f->rwlock);
1306                 else
1307                         fio_rwlock_write(f->rwlock);
1308         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1309                 fio_mutex_down(f->lock);
1310
1311         td->file_locks[f->fileno] = td->o.file_lock_mode;
1312 }
1313
1314 void unlock_file(struct thread_data *td, struct fio_file *f)
1315 {
1316         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1317                 return;
1318
1319         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1320                 fio_rwlock_unlock(f->rwlock);
1321         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1322                 fio_mutex_up(f->lock);
1323
1324         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1325 }
1326
1327 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1328 {
1329         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1330                 return;
1331         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1332                 unlock_file(td, f);
1333 }
1334
1335 static int recurse_dir(struct thread_data *td, const char *dirname)
1336 {
1337         struct dirent *dir;
1338         int ret = 0;
1339         DIR *D;
1340
1341         D = opendir(dirname);
1342         if (!D) {
1343                 char buf[FIO_VERROR_SIZE];
1344
1345                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1346                 td_verror(td, errno, buf);
1347                 return 1;
1348         }
1349
1350         while ((dir = readdir(D)) != NULL) {
1351                 char full_path[PATH_MAX];
1352                 struct stat sb;
1353
1354                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1355                         continue;
1356
1357                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1358
1359                 if (lstat(full_path, &sb) == -1) {
1360                         if (errno != ENOENT) {
1361                                 td_verror(td, errno, "stat");
1362                                 return 1;
1363                         }
1364                 }
1365
1366                 if (S_ISREG(sb.st_mode)) {
1367                         add_file(td, full_path, 0, 1);
1368                         continue;
1369                 }
1370                 if (!S_ISDIR(sb.st_mode))
1371                         continue;
1372
1373                 ret = recurse_dir(td, full_path);
1374                 if (ret)
1375                         break;
1376         }
1377
1378         closedir(D);
1379         return ret;
1380 }
1381
1382 int add_dir_files(struct thread_data *td, const char *path)
1383 {
1384         int ret = recurse_dir(td, path);
1385
1386         if (!ret)
1387                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1388
1389         return ret;
1390 }
1391
1392 void dup_files(struct thread_data *td, struct thread_data *org)
1393 {
1394         struct fio_file *f;
1395         unsigned int i;
1396
1397         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1398
1399         if (!org->files)
1400                 return;
1401
1402         td->files = malloc(org->files_index * sizeof(f));
1403
1404         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1405                 td->file_locks = malloc(org->files_index);
1406
1407         for_each_file(org, f, i) {
1408                 struct fio_file *__f;
1409
1410                 __f = smalloc(sizeof(*__f));
1411                 if (!__f) {
1412                         log_err("fio: smalloc OOM\n");
1413                         assert(0);
1414                 }
1415                 __f->fd = -1;
1416                 fio_file_reset(td, __f);
1417
1418                 if (f->file_name) {
1419                         __f->file_name = smalloc_strdup(f->file_name);
1420                         if (!__f->file_name) {
1421                                 log_err("fio: smalloc OOM\n");
1422                                 assert(0);
1423                         }
1424
1425                         __f->filetype = f->filetype;
1426                 }
1427
1428                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1429                         __f->lock = f->lock;
1430                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1431                         __f->rwlock = f->rwlock;
1432
1433                 td->files[i] = __f;
1434         }
1435 }
1436
1437 /*
1438  * Returns the index that matches the filename, or -1 if not there
1439  */
1440 int get_fileno(struct thread_data *td, const char *fname)
1441 {
1442         struct fio_file *f;
1443         unsigned int i;
1444
1445         for_each_file(td, f, i)
1446                 if (!strcmp(f->file_name, fname))
1447                         return i;
1448
1449         return -1;
1450 }
1451
1452 /*
1453  * For log usage, where we add/open/close files automatically
1454  */
1455 void free_release_files(struct thread_data *td)
1456 {
1457         close_files(td);
1458         td->files_index = 0;
1459         td->nr_normal_files = 0;
1460 }
1461
1462 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1463 {
1464         f->last_pos = f->file_offset;
1465         f->last_start = -1ULL;
1466         if (f->io_axmap)
1467                 axmap_reset(f->io_axmap);
1468         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1469                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1470 }
1471
1472 int fio_files_done(struct thread_data *td)
1473 {
1474         struct fio_file *f;
1475         unsigned int i;
1476
1477         for_each_file(td, f, i)
1478                 if (!fio_file_done(f))
1479                         return 0;
1480
1481         return 1;
1482 }
1483
1484 /* free memory used in initialization phase only */
1485 void filesetup_mem_free() {
1486         free_already_allocated();
1487 }