filesetup.c

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