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