filesetup.c

   1 #include <unistd.h>
   2 #include <fcntl.h>
   3 #include <string.h>
   4 #include <assert.h>
   5 #include <dirent.h>
   6 #include <libgen.h>
   7 #include <sys/stat.h>
   8 #include <sys/mman.h>
   9 #include <sys/types.h>
  10
  11 #include "fio.h"
  12 #include "smalloc.h"
  13 #include "filehash.h"
  14 #include "os/os.h"
  15 #include "hash.h"
  16 #include "lib/axmap.h"
  17
  18 #ifdef CONFIG_LINUX_FALLOCATE
  19 #include <linux/falloc.h>
  20 #endif
  21
  22 static int root_warn;
  23
  24 static inline void clear_error(struct thread_data *td)
  25 {
  26         td->error = 0;
  27         td->verror[0] = '\0';
  28 }
  29
  30 /*
  31  * Leaves f->fd open on success, caller must close
  32  */
  33 static int extend_file(struct thread_data *td, struct fio_file *f)
  34 {
  35         int r, new_layout = 0, unlink_file = 0, flags;
  36         unsigned long long left;
  37         unsigned int bs;
  38         char *b;
  39
  40         if (read_only) {
  41                 log_err("fio: refusing extend of file due to read-only\n");
  42                 return 0;
  43         }
  44
  45         /*
  46          * check if we need to lay the file out complete again. fio
  47          * does that for operations involving reads, or for writes
  48          * where overwrite is set
  49          */
  50         if (td_read(td) || (td_write(td) && td->o.overwrite) ||
  51             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
  52                 new_layout = 1;
  53         if (td_write(td) && !td->o.overwrite)
  54                 unlink_file = 1;
  55
  56         if (unlink_file || new_layout) {
  57                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  58                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
  59                         td_verror(td, errno, "unlink");
  60                         return 1;
  61                 }
  62         }
  63
  64         flags = O_WRONLY | O_CREAT;
  65         if (new_layout)
  66                 flags |= O_TRUNC;
  67
  68         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  69         f->fd = open(f->file_name, flags, 0644);
  70         if (f->fd < 0) {
  71                 td_verror(td, errno, "open");
  72                 return 1;
  73         }
  74
  75 #ifdef CONFIG_POSIX_FALLOCATE
  76         if (!td->o.fill_device) {
  77                 switch (td->o.fallocate_mode) {
  78                 case FIO_FALLOCATE_NONE:
  79                         break;
  80                 case FIO_FALLOCATE_POSIX:
  81                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  82                                  f->file_name,
  83                                  (unsigned long long) f->real_file_size);
  84
  85                         r = posix_fallocate(f->fd, 0, f->real_file_size);
  86                         if (r > 0) {
  87                                 log_err("fio: posix_fallocate fails: %s\n",
  88                                                 strerror(r));
  89                         }
  90                         break;
  91 #ifdef CONFIG_LINUX_FALLOCATE
  92                 case FIO_FALLOCATE_KEEP_SIZE:
  93                         dprint(FD_FILE,
  94                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
  95                                 "file %s size %llu\n", f->file_name,
  96                                 (unsigned long long) f->real_file_size);
  97
  98                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
  99                                         f->real_file_size);
 100                         if (r != 0)
 101                                 td_verror(td, errno, "fallocate");
 102
 103                         break;
 104 #endif /* CONFIG_LINUX_FALLOCATE */
 105                 default:
 106                         log_err("fio: unknown fallocate mode: %d\n",
 107                                 td->o.fallocate_mode);
 108                         assert(0);
 109                 }
 110         }
 111 #endif /* CONFIG_POSIX_FALLOCATE */
 112
 113         if (!new_layout)
 114                 goto done;
 115
 116         /*
 117          * The size will be -1ULL when fill_device is used, so don't truncate
 118          * or fallocate this file, just write it
 119          */
 120         if (!td->o.fill_device) {
 121                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 122                                         (unsigned long long) f->real_file_size);
 123                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 124                         if (errno != EFBIG) {
 125                                 td_verror(td, errno, "ftruncate");
 126                                 goto err;
 127                         }
 128                 }
 129         }
 130
 131         b = malloc(td->o.max_bs[DDIR_WRITE]);
 132
 133         left = f->real_file_size;
 134         while (left && !td->terminate) {
 135                 bs = td->o.max_bs[DDIR_WRITE];
 136                 if (bs > left)
 137                         bs = left;
 138
 139                 fill_io_buffer(td, b, bs, bs);
 140
 141                 r = write(f->fd, b, bs);
 142
 143                 if (r > 0) {
 144                         left -= r;
 145                         continue;
 146                 } else {
 147                         if (r < 0) {
 148                                 int __e = errno;
 149
 150                                 if (__e == ENOSPC) {
 151                                         if (td->o.fill_device)
 152                                                 break;
 153                                         log_info("fio: ENOSPC on laying out "
 154                                                  "file, stopping\n");
 155                                         break;
 156                                 }
 157                                 td_verror(td, errno, "write");
 158                         } else
 159                                 td_verror(td, EIO, "write");
 160
 161                         break;
 162                 }
 163         }
 164
 165         if (td->terminate) {
 166                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 167                 unlink(f->file_name);
 168         } else if (td->o.create_fsync) {
 169                 if (fsync(f->fd) < 0) {
 170                         td_verror(td, errno, "fsync");
 171                         goto err;
 172                 }
 173         }
 174         if (td->o.fill_device && !td_write(td)) {
 175                 fio_file_clear_size_known(f);
 176                 if (td_io_get_file_size(td, f))
 177                         goto err;
 178                 if (f->io_size > f->real_file_size)
 179                         f->io_size = f->real_file_size;
 180         }
 181
 182         free(b);
 183 done:
 184         return 0;
 185 err:
 186         close(f->fd);
 187         f->fd = -1;
 188         return 1;
 189 }
 190
 191 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 192 {
 193         int r, did_open = 0, old_runstate;
 194         unsigned long long left;
 195         unsigned int bs;
 196         char *b;
 197
 198         if (td->io_ops->flags & FIO_PIPEIO)
 199                 return 0;
 200
 201         if (!fio_file_open(f)) {
 202                 if (td->io_ops->open_file(td, f)) {
 203                         log_err("fio: cannot pre-read, failed to open file\n");
 204                         return 1;
 205                 }
 206                 did_open = 1;
 207         }
 208
 209         old_runstate = td->runstate;
 210         td_set_runstate(td, TD_PRE_READING);
 211
 212         bs = td->o.max_bs[DDIR_READ];
 213         b = malloc(bs);
 214         memset(b, 0, bs);
 215
 216         lseek(f->fd, f->file_offset, SEEK_SET);
 217         left = f->io_size;
 218
 219         while (left && !td->terminate) {
 220                 if (bs > left)
 221                         bs = left;
 222
 223                 r = read(f->fd, b, bs);
 224
 225                 if (r == (int) bs) {
 226                         left -= bs;
 227                         continue;
 228                 } else {
 229                         td_verror(td, EIO, "pre_read");
 230                         break;
 231                 }
 232         }
 233
 234         td_set_runstate(td, old_runstate);
 235
 236         if (did_open)
 237                 td->io_ops->close_file(td, f);
 238         free(b);
 239         return 0;
 240 }
 241
 242 static unsigned long long get_rand_file_size(struct thread_data *td)
 243 {
 244         unsigned long long ret, sized;
 245         unsigned long r;
 246
 247         if (td->o.use_os_rand) {
 248                 r = os_random_long(&td->file_size_state);
 249                 sized = td->o.file_size_high - td->o.file_size_low;
 250                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 251         } else {
 252                 r = __rand(&td->__file_size_state);
 253                 sized = td->o.file_size_high - td->o.file_size_low;
 254                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 255         }
 256
 257         ret += td->o.file_size_low;
 258         ret -= (ret % td->o.rw_min_bs);
 259         return ret;
 260 }
 261
 262 static int file_size(struct thread_data *td, struct fio_file *f)
 263 {
 264         struct stat st;
 265
 266         if (stat(f->file_name, &st) == -1) {
 267                 td_verror(td, errno, "fstat");
 268                 return 1;
 269         }
 270
 271         f->real_file_size = st.st_size;
 272         return 0;
 273 }
 274
 275 static int bdev_size(struct thread_data *td, struct fio_file *f)
 276 {
 277         unsigned long long bytes = 0;
 278         int r;
 279
 280         if (td->io_ops->open_file(td, f)) {
 281                 log_err("fio: failed opening blockdev %s for size check\n",
 282                         f->file_name);
 283                 return 1;
 284         }
 285
 286         r = blockdev_size(f, &bytes);
 287         if (r) {
 288                 td_verror(td, r, "blockdev_size");
 289                 goto err;
 290         }
 291
 292         if (!bytes) {
 293                 log_err("%s: zero sized block device?\n", f->file_name);
 294                 goto err;
 295         }
 296
 297         f->real_file_size = bytes;
 298         td->io_ops->close_file(td, f);
 299         return 0;
 300 err:
 301         td->io_ops->close_file(td, f);
 302         return 1;
 303 }
 304
 305 static int char_size(struct thread_data *td, struct fio_file *f)
 306 {
 307 #ifdef FIO_HAVE_CHARDEV_SIZE
 308         unsigned long long bytes = 0;
 309         int r;
 310
 311         if (td->io_ops->open_file(td, f)) {
 312                 log_err("fio: failed opening blockdev %s for size check\n",
 313                         f->file_name);
 314                 return 1;
 315         }
 316
 317         r = chardev_size(f, &bytes);
 318         if (r) {
 319                 td_verror(td, r, "chardev_size");
 320                 goto err;
 321         }
 322
 323         if (!bytes) {
 324                 log_err("%s: zero sized char device?\n", f->file_name);
 325                 goto err;
 326         }
 327
 328         f->real_file_size = bytes;
 329         td->io_ops->close_file(td, f);
 330         return 0;
 331 err:
 332         td->io_ops->close_file(td, f);
 333         return 1;
 334 #else
 335         f->real_file_size = -1ULL;
 336         return 0;
 337 #endif
 338 }
 339
 340 static int get_file_size(struct thread_data *td, struct fio_file *f)
 341 {
 342         int ret = 0;
 343
 344         if (fio_file_size_known(f))
 345                 return 0;
 346
 347         if (f->filetype == FIO_TYPE_FILE)
 348                 ret = file_size(td, f);
 349         else if (f->filetype == FIO_TYPE_BD)
 350                 ret = bdev_size(td, f);
 351         else if (f->filetype == FIO_TYPE_CHAR)
 352                 ret = char_size(td, f);
 353         else
 354                 f->real_file_size = -1;
 355
 356         if (ret)
 357                 return ret;
 358
 359         if (f->file_offset > f->real_file_size) {
 360                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 361                                         (unsigned long long) f->file_offset,
 362                                         (unsigned long long) f->real_file_size);
 363                 return 1;
 364         }
 365
 366         fio_file_set_size_known(f);
 367         return 0;
 368 }
 369
 370 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 371                                    unsigned long long off,
 372                                    unsigned long long len)
 373 {
 374         int ret = 0;
 375
 376         if (len == -1ULL)
 377                 len = f->io_size;
 378         if (off == -1ULL)
 379                 off = f->file_offset;
 380
 381         if (len == -1ULL || off == -1ULL)
 382                 return 0;
 383
 384         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 385                                                                 len);
 386
 387         /*
 388          * FIXME: add blockdev flushing too
 389          */
 390         if (f->mmap_ptr) {
 391                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 392 #ifdef FIO_MADV_FREE
 393                 if (f->filetype == FIO_TYPE_BD)
 394                         (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 395 #endif
 396         } else if (f->filetype == FIO_TYPE_FILE) {
 397                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 398         } else if (f->filetype == FIO_TYPE_BD) {
 399                 ret = blockdev_invalidate_cache(f);
 400                 if (ret < 0 && errno == EACCES && geteuid()) {
 401                         if (!root_warn) {
 402                                 log_err("fio: only root may flush block "
 403                                         "devices. Cache flush bypassed!\n");
 404                                 root_warn = 1;
 405                         }
 406                         ret = 0;
 407                 }
 408         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 409                 ret = 0;
 410
 411         if (ret < 0) {
 412                 td_verror(td, errno, "invalidate_cache");
 413                 return 1;
 414         } else if (ret > 0) {
 415                 td_verror(td, ret, "invalidate_cache");
 416                 return 1;
 417         }
 418
 419         return ret;
 420
 421 }
 422
 423 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 424 {
 425         if (!fio_file_open(f))
 426                 return 0;
 427
 428         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 429 }
 430
 431 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 432 {
 433         int ret = 0;
 434
 435         dprint(FD_FILE, "fd close %s\n", f->file_name);
 436
 437         remove_file_hash(f);
 438
 439         if (close(f->fd) < 0)
 440                 ret = errno;
 441
 442         f->fd = -1;
 443
 444         if (f->shadow_fd != -1) {
 445                 close(f->shadow_fd);
 446                 f->shadow_fd = -1;
 447         }
 448
 449         f->engine_data = 0;
 450         return ret;
 451 }
 452
 453 int file_lookup_open(struct fio_file *f, int flags)
 454 {
 455         struct fio_file *__f;
 456         int from_hash;
 457
 458         __f = lookup_file_hash(f->file_name);
 459         if (__f) {
 460                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 461                 /*
 462                  * racy, need the __f->lock locked
 463                  */
 464                 f->lock = __f->lock;
 465                 from_hash = 1;
 466         } else {
 467                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 468                 from_hash = 0;
 469         }
 470
 471         f->fd = open(f->file_name, flags, 0600);
 472         return from_hash;
 473 }
 474
 475 static int file_close_shadow_fds(struct thread_data *td)
 476 {
 477         struct fio_file *f;
 478         int num_closed = 0;
 479         unsigned int i;
 480
 481         for_each_file(td, f, i) {
 482                 if (f->shadow_fd == -1)
 483                         continue;
 484
 485                 close(f->shadow_fd);
 486                 f->shadow_fd = -1;
 487                 num_closed++;
 488         }
 489
 490         return num_closed;
 491 }
 492
 493 int generic_open_file(struct thread_data *td, struct fio_file *f)
 494 {
 495         int is_std = 0;
 496         int flags = 0;
 497         int from_hash = 0;
 498
 499         dprint(FD_FILE, "fd open %s\n", f->file_name);
 500
 501         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 502                 log_err("fio: trim only applies to block device\n");
 503                 return 1;
 504         }
 505
 506         if (!strcmp(f->file_name, "-")) {
 507                 if (td_rw(td)) {
 508                         log_err("fio: can't read/write to stdin/out\n");
 509                         return 1;
 510                 }
 511                 is_std = 1;
 512
 513                 /*
 514                  * move output logging to stderr, if we are writing to stdout
 515                  */
 516                 if (td_write(td))
 517                         f_out = stderr;
 518         }
 519
 520         if (td_trim(td))
 521                 goto skip_flags;
 522         if (td->o.odirect)
 523                 flags |= OS_O_DIRECT;
 524         if (td->o.oatomic) {
 525                 if (!FIO_O_ATOMIC) {
 526                         td_verror(td, EINVAL, "OS does not support atomic IO");
 527                         return 1;
 528                 }
 529                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 530         }
 531         if (td->o.sync_io)
 532                 flags |= O_SYNC;
 533         if (td->o.create_on_open)
 534                 flags |= O_CREAT;
 535 skip_flags:
 536         if (f->filetype != FIO_TYPE_FILE)
 537                 flags |= FIO_O_NOATIME;
 538
 539 open_again:
 540         if (td_write(td)) {
 541                 if (!read_only)
 542                         flags |= O_RDWR;
 543
 544                 if (f->filetype == FIO_TYPE_FILE)
 545                         flags |= O_CREAT;
 546
 547                 if (is_std)
 548                         f->fd = dup(STDOUT_FILENO);
 549                 else
 550                         from_hash = file_lookup_open(f, flags);
 551         } else if (td_read(td)) {
 552                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 553                         flags |= O_RDWR;
 554                 else
 555                         flags |= O_RDONLY;
 556
 557                 if (is_std)
 558                         f->fd = dup(STDIN_FILENO);
 559                 else
 560                         from_hash = file_lookup_open(f, flags);
 561         } else { //td trim
 562                 flags |= O_RDWR;
 563                 from_hash = file_lookup_open(f, flags);
 564         }
 565
 566         if (f->fd == -1) {
 567                 char buf[FIO_VERROR_SIZE];
 568                 int __e = errno;
 569
 570                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 571                         flags &= ~FIO_O_NOATIME;
 572                         goto open_again;
 573                 }
 574                 if (__e == EMFILE && file_close_shadow_fds(td))
 575                         goto open_again;
 576
 577                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 578
 579                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 580                         log_err("fio: looks like your file system does not " \
 581                                 "support direct=1/buffered=0\n");
 582                 }
 583
 584                 td_verror(td, __e, buf);
 585         }
 586
 587         if (!from_hash && f->fd != -1) {
 588                 if (add_file_hash(f)) {
 589                         int fio_unused ret;
 590
 591                         /*
 592                          * Stash away descriptor for later close. This is to
 593                          * work-around a "feature" on Linux, where a close of
 594                          * an fd that has been opened for write will trigger
 595                          * udev to call blkid to check partitions, fs id, etc.
 596                          * That pollutes the device cache, which can slow down
 597                          * unbuffered accesses.
 598                          */
 599                         if (f->shadow_fd == -1)
 600                                 f->shadow_fd = f->fd;
 601                         else {
 602                                 /*
 603                                  * OK to ignore, we haven't done anything
 604                                  * with it
 605                                  */
 606                                 ret = generic_close_file(td, f);
 607                         }
 608                         goto open_again;
 609                 }
 610         }
 611
 612         return 0;
 613 }
 614
 615 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 616 {
 617         return get_file_size(td, f);
 618 }
 619
 620 /*
 621  * open/close all files, so that ->real_file_size gets set
 622  */
 623 static int get_file_sizes(struct thread_data *td)
 624 {
 625         struct fio_file *f;
 626         unsigned int i;
 627         int err = 0;
 628
 629         for_each_file(td, f, i) {
 630                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 631                                                                 f->file_name);
 632
 633                 if (td_io_get_file_size(td, f)) {
 634                         if (td->error != ENOENT) {
 635                                 log_err("%s\n", td->verror);
 636                                 err = 1;
 637                         }
 638                         clear_error(td);
 639                 }
 640
 641                 if (f->real_file_size == -1ULL && td->o.size)
 642                         f->real_file_size = td->o.size / td->o.nr_files;
 643         }
 644
 645         return err;
 646 }
 647
 648 struct fio_mount {
 649         struct flist_head list;
 650         const char *base;
 651         char __base[256];
 652         unsigned int key;
 653 };
 654
 655 /*
 656  * Get free number of bytes for each file on each unique mount.
 657  */
 658 static unsigned long long get_fs_free_counts(struct thread_data *td)
 659 {
 660         struct flist_head *n, *tmp;
 661         unsigned long long ret = 0;
 662         struct fio_mount *fm;
 663         FLIST_HEAD(list);
 664         struct fio_file *f;
 665         unsigned int i;
 666
 667         for_each_file(td, f, i) {
 668                 struct stat sb;
 669                 char buf[256];
 670
 671                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 672                         if (f->real_file_size != -1ULL)
 673                                 ret += f->real_file_size;
 674                         continue;
 675                 } else if (f->filetype != FIO_TYPE_FILE)
 676                         continue;
 677
 678                 strcpy(buf, f->file_name);
 679
 680                 if (stat(buf, &sb) < 0) {
 681                         if (errno != ENOENT)
 682                                 break;
 683                         strcpy(buf, ".");
 684                         if (stat(buf, &sb) < 0)
 685                                 break;
 686                 }
 687
 688                 fm = NULL;
 689                 flist_for_each(n, &list) {
 690                         fm = flist_entry(n, struct fio_mount, list);
 691                         if (fm->key == sb.st_dev)
 692                                 break;
 693
 694                         fm = NULL;
 695                 }
 696
 697                 if (fm)
 698                         continue;
 699
 700                 fm = malloc(sizeof(*fm));
 701                 strcpy(fm->__base, buf);
 702                 fm->base = basename(fm->__base);
 703                 fm->key = sb.st_dev;
 704                 flist_add(&fm->list, &list);
 705         }
 706
 707         flist_for_each_safe(n, tmp, &list) {
 708                 unsigned long long sz;
 709
 710                 fm = flist_entry(n, struct fio_mount, list);
 711                 flist_del(&fm->list);
 712
 713                 sz = get_fs_size(fm->base);
 714                 if (sz && sz != -1ULL)
 715                         ret += sz;
 716
 717                 free(fm);
 718         }
 719
 720         return ret;
 721 }
 722
 723 uint64_t get_start_offset(struct thread_data *td)
 724 {
 725         return td->o.start_offset +
 726                 (td->thread_number - 1) * td->o.offset_increment;
 727 }
 728
 729 /*
 730  * Open the files and setup files sizes, creating files if necessary.
 731  */
 732 int setup_files(struct thread_data *td)
 733 {
 734         unsigned long long total_size, extend_size;
 735         struct thread_options *o = &td->o;
 736         struct fio_file *f;
 737         unsigned int i, nr_fs_extra = 0;
 738         int err = 0, need_extend;
 739         int old_state;
 740         const unsigned int bs = td_min_bs(td);
 741         uint64_t fs = 0;
 742
 743         dprint(FD_FILE, "setup files\n");
 744
 745         old_state = td->runstate;
 746         td_set_runstate(td, TD_SETTING_UP);
 747
 748         if (o->read_iolog_file)
 749                 goto done;
 750
 751         /*
 752          * if ioengine defines a setup() method, it's responsible for
 753          * opening the files and setting f->real_file_size to indicate
 754          * the valid range for that file.
 755          */
 756         if (td->io_ops->setup)
 757                 err = td->io_ops->setup(td);
 758         else
 759                 err = get_file_sizes(td);
 760
 761         if (err)
 762                 goto err_out;
 763
 764         /*
 765          * check sizes. if the files/devices do not exist and the size
 766          * isn't passed to fio, abort.
 767          */
 768         total_size = 0;
 769         for_each_file(td, f, i) {
 770                 if (f->real_file_size == -1ULL)
 771                         total_size = -1ULL;
 772                 else
 773                         total_size += f->real_file_size;
 774         }
 775
 776         if (o->fill_device)
 777                 td->fill_device_size = get_fs_free_counts(td);
 778
 779         /*
 780          * device/file sizes are zero and no size given, punt
 781          */
 782         if ((!total_size || total_size == -1ULL) && !o->size &&
 783             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 784             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 785                 log_err("%s: you need to specify size=\n", o->name);
 786                 td_verror(td, EINVAL, "total_file_size");
 787                 goto err_out;
 788         }
 789
 790         /*
 791          * Calculate per-file size and potential extra size for the
 792          * first files, if needed.
 793          */
 794         if (!o->file_size_low) {
 795                 uint64_t all_fs;
 796
 797                 fs = o->size / o->nr_files;
 798                 all_fs = fs * o->nr_files;
 799
 800                 if (all_fs < o->size)
 801                         nr_fs_extra = (o->size - all_fs) / bs;
 802         }
 803
 804         /*
 805          * now file sizes are known, so we can set ->io_size. if size= is
 806          * not given, ->io_size is just equal to ->real_file_size. if size
 807          * is given, ->io_size is size / nr_files.
 808          */
 809         extend_size = total_size = 0;
 810         need_extend = 0;
 811         for_each_file(td, f, i) {
 812                 f->file_offset = get_start_offset(td);
 813
 814                 if (!o->file_size_low) {
 815                         /*
 816                          * no file size range given, file size is equal to
 817                          * total size divided by number of files. If that is
 818                          * zero, set it to the real file size. If the size
 819                          * doesn't divide nicely with the min blocksize,
 820                          * make the first files bigger.
 821                          */
 822                         f->io_size = fs;
 823                         if (nr_fs_extra) {
 824                                 nr_fs_extra--;
 825                                 f->io_size += bs;
 826                         }
 827
 828                         if (!f->io_size)
 829                                 f->io_size = f->real_file_size - f->file_offset;
 830                 } else if (f->real_file_size < o->file_size_low ||
 831                            f->real_file_size > o->file_size_high) {
 832                         if (f->file_offset > o->file_size_low)
 833                                 goto err_offset;
 834                         /*
 835                          * file size given. if it's fixed, use that. if it's a
 836                          * range, generate a random size in-between.
 837                          */
 838                         if (o->file_size_low == o->file_size_high)
 839                                 f->io_size = o->file_size_low - f->file_offset;
 840                         else {
 841                                 f->io_size = get_rand_file_size(td)
 842                                                 - f->file_offset;
 843                         }
 844                 } else
 845                         f->io_size = f->real_file_size - f->file_offset;
 846
 847                 if (f->io_size == -1ULL)
 848                         total_size = -1ULL;
 849                 else {
 850                         if (o->size_percent)
 851                                 f->io_size = (f->io_size * o->size_percent) / 100;
 852                         total_size += f->io_size;
 853                 }
 854
 855                 if (f->filetype == FIO_TYPE_FILE &&
 856                     (f->io_size + f->file_offset) > f->real_file_size &&
 857                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 858                         if (!o->create_on_open) {
 859                                 need_extend++;
 860                                 extend_size += (f->io_size + f->file_offset);
 861                         } else
 862                                 f->real_file_size = f->io_size + f->file_offset;
 863                         fio_file_set_extend(f);
 864                 }
 865         }
 866
 867         if (!o->size || o->size > total_size)
 868                 o->size = total_size;
 869
 870         /*
 871          * See if we need to extend some files
 872          */
 873         if (need_extend) {
 874                 temp_stall_ts = 1;
 875                 if (output_format == FIO_OUTPUT_NORMAL)
 876                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 877                                  " %lluMB)\n", o->name, need_extend,
 878                                         extend_size >> 20);
 879
 880                 for_each_file(td, f, i) {
 881                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 882
 883                         if (!fio_file_extend(f))
 884                                 continue;
 885
 886                         assert(f->filetype == FIO_TYPE_FILE);
 887                         fio_file_clear_extend(f);
 888                         if (!o->fill_device) {
 889                                 old_len = f->real_file_size;
 890                                 extend_len = f->io_size + f->file_offset -
 891                                                 old_len;
 892                         }
 893                         f->real_file_size = (f->io_size + f->file_offset);
 894                         err = extend_file(td, f);
 895                         if (err)
 896                                 break;
 897
 898                         err = __file_invalidate_cache(td, f, old_len,
 899                                                                 extend_len);
 900                         close(f->fd);
 901                         f->fd = -1;
 902                         if (err)
 903                                 break;
 904                 }
 905                 temp_stall_ts = 0;
 906         }
 907
 908         if (err)
 909                 goto err_out;
 910
 911         if (!o->zone_size)
 912                 o->zone_size = o->size;
 913
 914         /*
 915          * iolog already set the total io size, if we read back
 916          * stored entries.
 917          */
 918         if (!o->read_iolog_file)
 919                 td->total_io_size = o->size * o->loops;
 920
 921 done:
 922         if (o->create_only)
 923                 td->done = 1;
 924
 925         td_set_runstate(td, old_state);
 926         return 0;
 927 err_offset:
 928         log_err("%s: you need to specify valid offset=\n", o->name);
 929 err_out:
 930         td_set_runstate(td, old_state);
 931         return 1;
 932 }
 933
 934 int pre_read_files(struct thread_data *td)
 935 {
 936         struct fio_file *f;
 937         unsigned int i;
 938
 939         dprint(FD_FILE, "pre_read files\n");
 940
 941         for_each_file(td, f, i) {
 942                 pre_read_file(td, f);
 943         }
 944
 945         return 1;
 946 }
 947
 948 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 949 {
 950         unsigned int range_size, seed;
 951         unsigned long nranges;
 952         uint64_t file_size;
 953
 954         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 955         file_size = min(f->real_file_size, f->io_size);
 956
 957         nranges = (file_size + range_size - 1) / range_size;
 958
 959         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 960         if (!td->o.rand_repeatable)
 961                 seed = td->rand_seeds[4];
 962
 963         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 964                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 965         else
 966                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 967
 968         return 1;
 969 }
 970
 971 static int init_rand_distribution(struct thread_data *td)
 972 {
 973         struct fio_file *f;
 974         unsigned int i;
 975         int state;
 976
 977         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 978                 return 0;
 979
 980         state = td->runstate;
 981         td_set_runstate(td, TD_SETTING_UP);
 982         for_each_file(td, f, i)
 983                 __init_rand_distribution(td, f);
 984         td_set_runstate(td, state);
 985
 986         return 1;
 987 }
 988
 989 int init_random_map(struct thread_data *td)
 990 {
 991         unsigned long long blocks;
 992         struct fio_file *f;
 993         unsigned int i;
 994
 995         if (init_rand_distribution(td))
 996                 return 0;
 997         if (!td_random(td))
 998                 return 0;
 999
1000         for_each_file(td, f, i) {
1001                 uint64_t file_size = min(f->real_file_size, f->io_size);
1002
1003                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1004
1005                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1006                         unsigned long seed;
1007
1008                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1009
1010                         if (!lfsr_init(&f->lfsr, blocks, seed, seed & 0xF))
1011                                 continue;
1012                 } else if (!td->o.norandommap) {
1013                         f->io_axmap = axmap_new(blocks);
1014                         if (f->io_axmap)
1015                                 continue;
1016                 } else if (td->o.norandommap)
1017                         continue;
1018
1019                 if (!td->o.softrandommap) {
1020                         log_err("fio: failed allocating random map. If running"
1021                                 " a large number of jobs, try the 'norandommap'"
1022                                 " option or set 'softrandommap'. Or give"
1023                                 " a larger --alloc-size to fio.\n");
1024                         return 1;
1025                 }
1026
1027                 log_info("fio: file %s failed allocating random map. Running "
1028                          "job without.\n", f->file_name);
1029         }
1030
1031         return 0;
1032 }
1033
1034 void close_files(struct thread_data *td)
1035 {
1036         struct fio_file *f;
1037         unsigned int i;
1038
1039         for_each_file(td, f, i) {
1040                 if (fio_file_open(f))
1041                         td_io_close_file(td, f);
1042         }
1043 }
1044
1045 void close_and_free_files(struct thread_data *td)
1046 {
1047         struct fio_file *f;
1048         unsigned int i;
1049
1050         dprint(FD_FILE, "close files\n");
1051
1052         for_each_file(td, f, i) {
1053                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1054                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1055                         unlink(f->file_name);
1056                 }
1057
1058                 if (fio_file_open(f))
1059                         td_io_close_file(td, f);
1060
1061                 remove_file_hash(f);
1062
1063                 sfree(f->file_name);
1064                 f->file_name = NULL;
1065                 axmap_free(f->io_axmap);
1066                 f->io_axmap = NULL;
1067                 sfree(f);
1068         }
1069
1070         td->o.filename = NULL;
1071         free(td->files);
1072         free(td->file_locks);
1073         td->files_index = 0;
1074         td->files = NULL;
1075         td->file_locks = NULL;
1076         td->o.file_lock_mode = FILE_LOCK_NONE;
1077         td->o.nr_files = 0;
1078 }
1079
1080 static void get_file_type(struct fio_file *f)
1081 {
1082         struct stat sb;
1083
1084         if (!strcmp(f->file_name, "-"))
1085                 f->filetype = FIO_TYPE_PIPE;
1086         else
1087                 f->filetype = FIO_TYPE_FILE;
1088
1089         /* \\.\ is the device namespace in Windows, where every file is
1090          * a block device */
1091         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1092                 f->filetype = FIO_TYPE_BD;
1093
1094         if (!stat(f->file_name, &sb)) {
1095                 if (S_ISBLK(sb.st_mode))
1096                         f->filetype = FIO_TYPE_BD;
1097                 else if (S_ISCHR(sb.st_mode))
1098                         f->filetype = FIO_TYPE_CHAR;
1099                 else if (S_ISFIFO(sb.st_mode))
1100                         f->filetype = FIO_TYPE_PIPE;
1101         }
1102 }
1103
1104 int add_file(struct thread_data *td, const char *fname)
1105 {
1106         int cur_files = td->files_index;
1107         char file_name[PATH_MAX];
1108         struct fio_file *f;
1109         int len = 0;
1110
1111         dprint(FD_FILE, "add file %s\n", fname);
1112
1113         f = smalloc(sizeof(*f));
1114         if (!f) {
1115                 log_err("fio: smalloc OOM\n");
1116                 assert(0);
1117         }
1118
1119         f->fd = -1;
1120         f->shadow_fd = -1;
1121         fio_file_reset(td, f);
1122
1123         if (td->files_size <= td->files_index) {
1124                 unsigned int new_size = td->o.nr_files + 1;
1125
1126                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1127
1128                 td->files = realloc(td->files, new_size * sizeof(f));
1129                 if (td->files == NULL) {
1130                         log_err("fio: realloc OOM\n");
1131                         assert(0);
1132                 }
1133                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1134                         td->file_locks = realloc(td->file_locks, new_size);
1135                         if (!td->file_locks) {
1136                                 log_err("fio: realloc OOM\n");
1137                                 assert(0);
1138                         }
1139                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1140                 }
1141                 td->files_size = new_size;
1142         }
1143         td->files[cur_files] = f;
1144         f->fileno = cur_files;
1145
1146         /*
1147          * init function, io engine may not be loaded yet
1148          */
1149         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1150                 f->real_file_size = -1ULL;
1151
1152         if (td->o.directory)
1153                 len = sprintf(file_name, "%s/", td->o.directory);
1154
1155         sprintf(file_name + len, "%s", fname);
1156         f->file_name = smalloc_strdup(file_name);
1157         if (!f->file_name) {
1158                 log_err("fio: smalloc OOM\n");
1159                 assert(0);
1160         }
1161
1162         get_file_type(f);
1163
1164         switch (td->o.file_lock_mode) {
1165         case FILE_LOCK_NONE:
1166                 break;
1167         case FILE_LOCK_READWRITE:
1168                 f->rwlock = fio_rwlock_init();
1169                 break;
1170         case FILE_LOCK_EXCLUSIVE:
1171                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1172                 break;
1173         default:
1174                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1175                 assert(0);
1176         }
1177
1178         td->files_index++;
1179         if (f->filetype == FIO_TYPE_FILE)
1180                 td->nr_normal_files++;
1181
1182         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1183                                                         cur_files);
1184
1185         return cur_files;
1186 }
1187
1188 int add_file_exclusive(struct thread_data *td, const char *fname)
1189 {
1190         struct fio_file *f;
1191         unsigned int i;
1192
1193         for_each_file(td, f, i) {
1194                 if (!strcmp(f->file_name, fname))
1195                         return i;
1196         }
1197
1198         return add_file(td, fname);
1199 }
1200
1201 void get_file(struct fio_file *f)
1202 {
1203         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1204         assert(fio_file_open(f));
1205         f->references++;
1206 }
1207
1208 int put_file(struct thread_data *td, struct fio_file *f)
1209 {
1210         int f_ret = 0, ret = 0;
1211
1212         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1213
1214         if (!fio_file_open(f)) {
1215                 assert(f->fd == -1);
1216                 return 0;
1217         }
1218
1219         assert(f->references);
1220         if (--f->references)
1221                 return 0;
1222
1223         if (should_fsync(td) && td->o.fsync_on_close)
1224                 f_ret = fsync(f->fd);
1225
1226         if (td->io_ops->close_file)
1227                 ret = td->io_ops->close_file(td, f);
1228
1229         if (!ret)
1230                 ret = f_ret;
1231
1232         td->nr_open_files--;
1233         fio_file_clear_open(f);
1234         assert(f->fd == -1);
1235         return ret;
1236 }
1237
1238 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1239 {
1240         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1241                 return;
1242
1243         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1244                 if (ddir == DDIR_READ)
1245                         fio_rwlock_read(f->rwlock);
1246                 else
1247                         fio_rwlock_write(f->rwlock);
1248         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1249                 fio_mutex_down(f->lock);
1250
1251         td->file_locks[f->fileno] = td->o.file_lock_mode;
1252 }
1253
1254 void unlock_file(struct thread_data *td, struct fio_file *f)
1255 {
1256         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1257                 return;
1258
1259         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1260                 fio_rwlock_unlock(f->rwlock);
1261         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1262                 fio_mutex_up(f->lock);
1263
1264         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1265 }
1266
1267 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1268 {
1269         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1270                 return;
1271         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1272                 unlock_file(td, f);
1273 }
1274
1275 static int recurse_dir(struct thread_data *td, const char *dirname)
1276 {
1277         struct dirent *dir;
1278         int ret = 0;
1279         DIR *D;
1280
1281         D = opendir(dirname);
1282         if (!D) {
1283                 char buf[FIO_VERROR_SIZE];
1284
1285                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1286                 td_verror(td, errno, buf);
1287                 return 1;
1288         }
1289
1290         while ((dir = readdir(D)) != NULL) {
1291                 char full_path[PATH_MAX];
1292                 struct stat sb;
1293
1294                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1295                         continue;
1296
1297                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1298
1299                 if (lstat(full_path, &sb) == -1) {
1300                         if (errno != ENOENT) {
1301                                 td_verror(td, errno, "stat");
1302                                 return 1;
1303                         }
1304                 }
1305
1306                 if (S_ISREG(sb.st_mode)) {
1307                         add_file(td, full_path);
1308                         td->o.nr_files++;
1309                         continue;
1310                 }
1311                 if (!S_ISDIR(sb.st_mode))
1312                         continue;
1313
1314                 ret = recurse_dir(td, full_path);
1315                 if (ret)
1316                         break;
1317         }
1318
1319         closedir(D);
1320         return ret;
1321 }
1322
1323 int add_dir_files(struct thread_data *td, const char *path)
1324 {
1325         int ret = recurse_dir(td, path);
1326
1327         if (!ret)
1328                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1329
1330         return ret;
1331 }
1332
1333 void dup_files(struct thread_data *td, struct thread_data *org)
1334 {
1335         struct fio_file *f;
1336         unsigned int i;
1337
1338         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1339
1340         if (!org->files)
1341                 return;
1342
1343         td->files = malloc(org->files_index * sizeof(f));
1344
1345         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1346                 td->file_locks = malloc(org->files_index);
1347
1348         for_each_file(org, f, i) {
1349                 struct fio_file *__f;
1350
1351                 __f = smalloc(sizeof(*__f));
1352                 if (!__f) {
1353                         log_err("fio: smalloc OOM\n");
1354                         assert(0);
1355                 }
1356                 __f->fd = -1;
1357                 fio_file_reset(td, __f);
1358
1359                 if (f->file_name) {
1360                         __f->file_name = smalloc_strdup(f->file_name);
1361                         if (!__f->file_name) {
1362                                 log_err("fio: smalloc OOM\n");
1363                                 assert(0);
1364                         }
1365
1366                         __f->filetype = f->filetype;
1367                 }
1368
1369                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1370                         __f->lock = f->lock;
1371                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1372                         __f->rwlock = f->rwlock;
1373
1374                 td->files[i] = __f;
1375         }
1376 }
1377
1378 /*
1379  * Returns the index that matches the filename, or -1 if not there
1380  */
1381 int get_fileno(struct thread_data *td, const char *fname)
1382 {
1383         struct fio_file *f;
1384         unsigned int i;
1385
1386         for_each_file(td, f, i)
1387                 if (!strcmp(f->file_name, fname))
1388                         return i;
1389
1390         return -1;
1391 }
1392
1393 /*
1394  * For log usage, where we add/open/close files automatically
1395  */
1396 void free_release_files(struct thread_data *td)
1397 {
1398         close_files(td);
1399         td->files_index = 0;
1400         td->nr_normal_files = 0;
1401 }
1402
1403 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1404 {
1405         f->last_pos = f->file_offset;
1406         f->last_start = -1ULL;
1407         if (f->io_axmap)
1408                 axmap_reset(f->io_axmap);
1409         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1410                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1411 }
1412
1413 int fio_files_done(struct thread_data *td)
1414 {
1415         struct fio_file *f;
1416         unsigned int i;
1417
1418         for_each_file(td, f, i)
1419                 if (!fio_file_done(f))
1420                         return 0;
1421
1422         return 1;
1423 }