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                 from_hash = 1;
 460         } else {
 461                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 462                 from_hash = 0;
 463         }
 464
 465         f->fd = open(f->file_name, flags, 0600);
 466         return from_hash;
 467 }
 468
 469 static int file_close_shadow_fds(struct thread_data *td)
 470 {
 471         struct fio_file *f;
 472         int num_closed = 0;
 473         unsigned int i;
 474
 475         for_each_file(td, f, i) {
 476                 if (f->shadow_fd == -1)
 477                         continue;
 478
 479                 close(f->shadow_fd);
 480                 f->shadow_fd = -1;
 481                 num_closed++;
 482         }
 483
 484         return num_closed;
 485 }
 486
 487 int generic_open_file(struct thread_data *td, struct fio_file *f)
 488 {
 489         int is_std = 0;
 490         int flags = 0;
 491         int from_hash = 0;
 492
 493         dprint(FD_FILE, "fd open %s\n", f->file_name);
 494
 495         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 496                 log_err("fio: trim only applies to block device\n");
 497                 return 1;
 498         }
 499
 500         if (!strcmp(f->file_name, "-")) {
 501                 if (td_rw(td)) {
 502                         log_err("fio: can't read/write to stdin/out\n");
 503                         return 1;
 504                 }
 505                 is_std = 1;
 506
 507                 /*
 508                  * move output logging to stderr, if we are writing to stdout
 509                  */
 510                 if (td_write(td))
 511                         f_out = stderr;
 512         }
 513
 514         if (td_trim(td))
 515                 goto skip_flags;
 516         if (td->o.odirect)
 517                 flags |= OS_O_DIRECT;
 518         if (td->o.sync_io)
 519                 flags |= O_SYNC;
 520         if (td->o.create_on_open)
 521                 flags |= O_CREAT;
 522 skip_flags:
 523         if (f->filetype != FIO_TYPE_FILE)
 524                 flags |= FIO_O_NOATIME;
 525
 526 open_again:
 527         if (td_write(td)) {
 528                 if (!read_only)
 529                         flags |= O_RDWR;
 530
 531                 if (f->filetype == FIO_TYPE_FILE)
 532                         flags |= O_CREAT;
 533
 534                 if (is_std)
 535                         f->fd = dup(STDOUT_FILENO);
 536                 else
 537                         from_hash = file_lookup_open(f, flags);
 538         } else if (td_read(td)) {
 539                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 540                         flags |= O_RDWR;
 541                 else
 542                         flags |= O_RDONLY;
 543
 544                 if (is_std)
 545                         f->fd = dup(STDIN_FILENO);
 546                 else
 547                         from_hash = file_lookup_open(f, flags);
 548         } else { //td trim
 549                 flags |= O_RDWR;
 550                 from_hash = file_lookup_open(f, flags);
 551         }
 552
 553         if (f->fd == -1) {
 554                 char buf[FIO_VERROR_SIZE];
 555                 int __e = errno;
 556
 557                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 558                         flags &= ~FIO_O_NOATIME;
 559                         goto open_again;
 560                 }
 561                 if (__e == EMFILE && file_close_shadow_fds(td))
 562                         goto open_again;
 563
 564                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 565
 566                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 567                         log_err("fio: looks like your file system does not " \
 568                                 "support direct=1/buffered=0\n");
 569                 }
 570
 571                 td_verror(td, __e, buf);
 572         }
 573
 574         if (!from_hash && f->fd != -1) {
 575                 if (add_file_hash(f)) {
 576                         int fio_unused ret;
 577
 578                         /*
 579                          * Stash away descriptor for later close. This is to
 580                          * work-around a "feature" on Linux, where a close of
 581                          * an fd that has been opened for write will trigger
 582                          * udev to call blkid to check partitions, fs id, etc.
 583                          * That polutes the device cache, which can slow down
 584                          * unbuffered accesses.
 585                          */
 586                         if (f->shadow_fd == -1)
 587                                 f->shadow_fd = f->fd;
 588                         else {
 589                                 /*
 590                                  * OK to ignore, we haven't done anything
 591                                  * with it
 592                                  */
 593                                 ret = generic_close_file(td, f);
 594                         }
 595                         goto open_again;
 596                 }
 597         }
 598
 599         return 0;
 600 }
 601
 602 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 603 {
 604         return get_file_size(td, f);
 605 }
 606
 607 /*
 608  * open/close all files, so that ->real_file_size gets set
 609  */
 610 static int get_file_sizes(struct thread_data *td)
 611 {
 612         struct fio_file *f;
 613         unsigned int i;
 614         int err = 0;
 615
 616         for_each_file(td, f, i) {
 617                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 618                                                                 f->file_name);
 619
 620                 if (td_io_get_file_size(td, f)) {
 621                         if (td->error != ENOENT) {
 622                                 log_err("%s\n", td->verror);
 623                                 err = 1;
 624                         }
 625                         clear_error(td);
 626                 }
 627
 628                 if (f->real_file_size == -1ULL && td->o.size)
 629                         f->real_file_size = td->o.size / td->o.nr_files;
 630         }
 631
 632         return err;
 633 }
 634
 635 struct fio_mount {
 636         struct flist_head list;
 637         const char *base;
 638         char __base[256];
 639         unsigned int key;
 640 };
 641
 642 /*
 643  * Get free number of bytes for each file on each unique mount.
 644  */
 645 static unsigned long long get_fs_free_counts(struct thread_data *td)
 646 {
 647         struct flist_head *n, *tmp;
 648         unsigned long long ret = 0;
 649         struct fio_mount *fm;
 650         FLIST_HEAD(list);
 651         struct fio_file *f;
 652         unsigned int i;
 653
 654         for_each_file(td, f, i) {
 655                 struct stat sb;
 656                 char buf[256];
 657
 658                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 659                         if (f->real_file_size != -1ULL)
 660                                 ret += f->real_file_size;
 661                         continue;
 662                 } else if (f->filetype != FIO_TYPE_FILE)
 663                         continue;
 664
 665                 strcpy(buf, f->file_name);
 666
 667                 if (stat(buf, &sb) < 0) {
 668                         if (errno != ENOENT)
 669                                 break;
 670                         strcpy(buf, ".");
 671                         if (stat(buf, &sb) < 0)
 672                                 break;
 673                 }
 674
 675                 fm = NULL;
 676                 flist_for_each(n, &list) {
 677                         fm = flist_entry(n, struct fio_mount, list);
 678                         if (fm->key == sb.st_dev)
 679                                 break;
 680
 681                         fm = NULL;
 682                 }
 683
 684                 if (fm)
 685                         continue;
 686
 687                 fm = malloc(sizeof(*fm));
 688                 strcpy(fm->__base, buf);
 689                 fm->base = basename(fm->__base);
 690                 fm->key = sb.st_dev;
 691                 flist_add(&fm->list, &list);
 692         }
 693
 694         flist_for_each_safe(n, tmp, &list) {
 695                 unsigned long long sz;
 696
 697                 fm = flist_entry(n, struct fio_mount, list);
 698                 flist_del(&fm->list);
 699
 700                 sz = get_fs_size(fm->base);
 701                 if (sz && sz != -1ULL)
 702                         ret += sz;
 703
 704                 free(fm);
 705         }
 706
 707         return ret;
 708 }
 709
 710 uint64_t get_start_offset(struct thread_data *td)
 711 {
 712         return td->o.start_offset +
 713                 (td->thread_number - 1) * td->o.offset_increment;
 714 }
 715
 716 /*
 717  * Open the files and setup files sizes, creating files if necessary.
 718  */
 719 int setup_files(struct thread_data *td)
 720 {
 721         unsigned long long total_size, extend_size;
 722         struct thread_options *o = &td->o;
 723         struct fio_file *f;
 724         unsigned int i;
 725         int err = 0, need_extend;
 726         int old_state;
 727
 728         dprint(FD_FILE, "setup files\n");
 729
 730         old_state = td->runstate;
 731         td_set_runstate(td, TD_SETTING_UP);
 732
 733         if (o->read_iolog_file)
 734                 goto done;
 735
 736         /*
 737          * if ioengine defines a setup() method, it's responsible for
 738          * opening the files and setting f->real_file_size to indicate
 739          * the valid range for that file.
 740          */
 741         if (td->io_ops->setup)
 742                 err = td->io_ops->setup(td);
 743         else
 744                 err = get_file_sizes(td);
 745
 746         if (err)
 747                 goto err_out;
 748
 749         /*
 750          * check sizes. if the files/devices do not exist and the size
 751          * isn't passed to fio, abort.
 752          */
 753         total_size = 0;
 754         for_each_file(td, f, i) {
 755                 if (f->real_file_size == -1ULL)
 756                         total_size = -1ULL;
 757                 else
 758                         total_size += f->real_file_size;
 759         }
 760
 761         if (o->fill_device)
 762                 td->fill_device_size = get_fs_free_counts(td);
 763
 764         /*
 765          * device/file sizes are zero and no size given, punt
 766          */
 767         if ((!total_size || total_size == -1ULL) && !o->size &&
 768             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 769             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 770                 log_err("%s: you need to specify size=\n", o->name);
 771                 td_verror(td, EINVAL, "total_file_size");
 772                 goto err_out;
 773         }
 774
 775         /*
 776          * now file sizes are known, so we can set ->io_size. if size= is
 777          * not given, ->io_size is just equal to ->real_file_size. if size
 778          * is given, ->io_size is size / nr_files.
 779          */
 780         extend_size = total_size = 0;
 781         need_extend = 0;
 782         for_each_file(td, f, i) {
 783                 f->file_offset = get_start_offset(td);
 784
 785                 if (!o->file_size_low) {
 786                         /*
 787                          * no file size range given, file size is equal to
 788                          * total size divided by number of files. if that is
 789                          * zero, set it to the real file size.
 790                          */
 791                         f->io_size = o->size / o->nr_files;
 792                         if (!f->io_size)
 793                                 f->io_size = f->real_file_size - f->file_offset;
 794                 } else if (f->real_file_size < o->file_size_low ||
 795                            f->real_file_size > o->file_size_high) {
 796                         if (f->file_offset > o->file_size_low)
 797                                 goto err_offset;
 798                         /*
 799                          * file size given. if it's fixed, use that. if it's a
 800                          * range, generate a random size in-between.
 801                          */
 802                         if (o->file_size_low == o->file_size_high)
 803                                 f->io_size = o->file_size_low - f->file_offset;
 804                         else {
 805                                 f->io_size = get_rand_file_size(td)
 806                                                 - f->file_offset;
 807                         }
 808                 } else
 809                         f->io_size = f->real_file_size - f->file_offset;
 810
 811                 if (f->io_size == -1ULL)
 812                         total_size = -1ULL;
 813                 else {
 814                         if (o->size_percent)
 815                                 f->io_size = (f->io_size * o->size_percent) / 100;
 816                         total_size += f->io_size;
 817                 }
 818
 819                 if (f->filetype == FIO_TYPE_FILE &&
 820                     (f->io_size + f->file_offset) > f->real_file_size &&
 821                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 822                         if (!o->create_on_open) {
 823                                 need_extend++;
 824                                 extend_size += (f->io_size + f->file_offset);
 825                         } else
 826                                 f->real_file_size = f->io_size + f->file_offset;
 827                         fio_file_set_extend(f);
 828                 }
 829         }
 830
 831         if (!o->size || o->size > total_size)
 832                 o->size = total_size;
 833
 834         /*
 835          * See if we need to extend some files
 836          */
 837         if (need_extend) {
 838                 temp_stall_ts = 1;
 839                 if (output_format == FIO_OUTPUT_NORMAL)
 840                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 841                                  " %lluMB)\n", o->name, need_extend,
 842                                         extend_size >> 20);
 843
 844                 for_each_file(td, f, i) {
 845                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 846
 847                         if (!fio_file_extend(f))
 848                                 continue;
 849
 850                         assert(f->filetype == FIO_TYPE_FILE);
 851                         fio_file_clear_extend(f);
 852                         if (!o->fill_device) {
 853                                 old_len = f->real_file_size;
 854                                 extend_len = f->io_size + f->file_offset -
 855                                                 old_len;
 856                         }
 857                         f->real_file_size = (f->io_size + f->file_offset);
 858                         err = extend_file(td, f);
 859                         if (err)
 860                                 break;
 861
 862                         err = __file_invalidate_cache(td, f, old_len,
 863                                                                 extend_len);
 864                         close(f->fd);
 865                         f->fd = -1;
 866                         if (err)
 867                                 break;
 868                 }
 869                 temp_stall_ts = 0;
 870         }
 871
 872         if (err)
 873                 goto err_out;
 874
 875         if (!o->zone_size)
 876                 o->zone_size = o->size;
 877
 878         /*
 879          * iolog already set the total io size, if we read back
 880          * stored entries.
 881          */
 882         if (!o->read_iolog_file)
 883                 td->total_io_size = o->size * o->loops;
 884
 885 done:
 886         if (o->create_only)
 887                 td->done = 1;
 888
 889         td_set_runstate(td, old_state);
 890         return 0;
 891 err_offset:
 892         log_err("%s: you need to specify valid offset=\n", o->name);
 893 err_out:
 894         td_set_runstate(td, old_state);
 895         return 1;
 896 }
 897
 898 int pre_read_files(struct thread_data *td)
 899 {
 900         struct fio_file *f;
 901         unsigned int i;
 902
 903         dprint(FD_FILE, "pre_read files\n");
 904
 905         for_each_file(td, f, i) {
 906                 pre_read_file(td, f);
 907         }
 908
 909         return 1;
 910 }
 911
 912 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 913 {
 914         unsigned int range_size, seed;
 915         unsigned long nranges;
 916         uint64_t file_size;
 917
 918         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 919         file_size = min(f->real_file_size, f->io_size);
 920
 921         nranges = (file_size + range_size - 1) / range_size;
 922
 923         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 924         if (!td->o.rand_repeatable)
 925                 seed = td->rand_seeds[4];
 926
 927         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 928                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 929         else
 930                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 931
 932         return 1;
 933 }
 934
 935 static int init_rand_distribution(struct thread_data *td)
 936 {
 937         struct fio_file *f;
 938         unsigned int i;
 939         int state;
 940
 941         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 942                 return 0;
 943
 944         state = td->runstate;
 945         td_set_runstate(td, TD_SETTING_UP);
 946         for_each_file(td, f, i)
 947                 __init_rand_distribution(td, f);
 948         td_set_runstate(td, state);
 949
 950         return 1;
 951 }
 952
 953 int init_random_map(struct thread_data *td)
 954 {
 955         unsigned long long blocks;
 956         struct fio_file *f;
 957         unsigned int i;
 958
 959         if (init_rand_distribution(td))
 960                 return 0;
 961         if (!td_random(td))
 962                 return 0;
 963
 964         for_each_file(td, f, i) {
 965                 uint64_t file_size = min(f->real_file_size, f->io_size);
 966
 967                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
 968
 969                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
 970                         unsigned long seed;
 971
 972                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
 973
 974                         if (!lfsr_init(&f->lfsr, blocks, seed, seed & 0xF))
 975                                 continue;
 976                 } else if (!td->o.norandommap) {
 977                         f->io_axmap = axmap_new(blocks);
 978                         if (f->io_axmap)
 979                                 continue;
 980                 } else if (td->o.norandommap)
 981                         continue;
 982
 983                 if (!td->o.softrandommap) {
 984                         log_err("fio: failed allocating random map. If running"
 985                                 " a large number of jobs, try the 'norandommap'"
 986                                 " option or set 'softrandommap'. Or give"
 987                                 " a larger --alloc-size to fio.\n");
 988                         return 1;
 989                 }
 990
 991                 log_info("fio: file %s failed allocating random map. Running "
 992                          "job without.\n", f->file_name);
 993         }
 994
 995         return 0;
 996 }
 997
 998 void close_files(struct thread_data *td)
 999 {
1000         struct fio_file *f;
1001         unsigned int i;
1002
1003         for_each_file(td, f, i) {
1004                 if (fio_file_open(f))
1005                         td_io_close_file(td, f);
1006         }
1007 }
1008
1009 void close_and_free_files(struct thread_data *td)
1010 {
1011         struct fio_file *f;
1012         unsigned int i;
1013
1014         dprint(FD_FILE, "close files\n");
1015
1016         for_each_file(td, f, i) {
1017                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1018                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1019                         unlink(f->file_name);
1020                 }
1021
1022                 if (fio_file_open(f))
1023                         td_io_close_file(td, f);
1024
1025                 remove_file_hash(f);
1026
1027                 sfree(f->file_name);
1028                 f->file_name = NULL;
1029                 axmap_free(f->io_axmap);
1030                 f->io_axmap = NULL;
1031                 sfree(f);
1032         }
1033
1034         td->o.filename = NULL;
1035         free(td->files);
1036         free(td->file_locks);
1037         td->files_index = 0;
1038         td->files = NULL;
1039         td->file_locks = NULL;
1040         td->o.nr_files = 0;
1041 }
1042
1043 static void get_file_type(struct fio_file *f)
1044 {
1045         struct stat sb;
1046
1047         if (!strcmp(f->file_name, "-"))
1048                 f->filetype = FIO_TYPE_PIPE;
1049         else
1050                 f->filetype = FIO_TYPE_FILE;
1051
1052         /* \\.\ is the device namespace in Windows, where every file is
1053          * a block device */
1054         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1055                 f->filetype = FIO_TYPE_BD;
1056
1057         if (!stat(f->file_name, &sb)) {
1058                 if (S_ISBLK(sb.st_mode))
1059                         f->filetype = FIO_TYPE_BD;
1060                 else if (S_ISCHR(sb.st_mode))
1061                         f->filetype = FIO_TYPE_CHAR;
1062                 else if (S_ISFIFO(sb.st_mode))
1063                         f->filetype = FIO_TYPE_PIPE;
1064         }
1065 }
1066
1067 int add_file(struct thread_data *td, const char *fname)
1068 {
1069         int cur_files = td->files_index;
1070         char file_name[PATH_MAX];
1071         struct fio_file *f;
1072         int len = 0;
1073
1074         dprint(FD_FILE, "add file %s\n", fname);
1075
1076         f = smalloc(sizeof(*f));
1077         if (!f) {
1078                 log_err("fio: smalloc OOM\n");
1079                 assert(0);
1080         }
1081
1082         f->fd = -1;
1083         f->shadow_fd = -1;
1084         fio_file_reset(td, f);
1085
1086         if (td->files_size <= td->files_index) {
1087                 unsigned int new_size = td->o.nr_files + 1;
1088
1089                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1090
1091                 td->files = realloc(td->files, new_size * sizeof(f));
1092                 if (td->files == NULL) {
1093                         log_err("fio: realloc OOM\n");
1094                         assert(0);
1095                 }
1096                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1097                         td->file_locks = realloc(td->file_locks, new_size);
1098                         if (!td->file_locks) {
1099                                 log_err("fio: realloc OOM\n");
1100                                 assert(0);
1101                         }
1102                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1103                 }
1104                 td->files_size = new_size;
1105         }
1106         td->files[cur_files] = f;
1107         f->fileno = cur_files;
1108
1109         /*
1110          * init function, io engine may not be loaded yet
1111          */
1112         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1113                 f->real_file_size = -1ULL;
1114
1115         if (td->o.directory)
1116                 len = sprintf(file_name, "%s/", td->o.directory);
1117
1118         sprintf(file_name + len, "%s", fname);
1119         f->file_name = smalloc_strdup(file_name);
1120         if (!f->file_name) {
1121                 log_err("fio: smalloc OOM\n");
1122                 assert(0);
1123         }
1124
1125         get_file_type(f);
1126
1127         switch (td->o.file_lock_mode) {
1128         case FILE_LOCK_NONE:
1129                 break;
1130         case FILE_LOCK_READWRITE:
1131                 f->rwlock = fio_rwlock_init();
1132                 break;
1133         case FILE_LOCK_EXCLUSIVE:
1134                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1135                 break;
1136         default:
1137                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1138                 assert(0);
1139         }
1140
1141         td->files_index++;
1142         if (f->filetype == FIO_TYPE_FILE)
1143                 td->nr_normal_files++;
1144
1145         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1146                                                         cur_files);
1147
1148         return cur_files;
1149 }
1150
1151 int add_file_exclusive(struct thread_data *td, const char *fname)
1152 {
1153         struct fio_file *f;
1154         unsigned int i;
1155
1156         for_each_file(td, f, i) {
1157                 if (!strcmp(f->file_name, fname))
1158                         return i;
1159         }
1160
1161         return add_file(td, fname);
1162 }
1163
1164 void get_file(struct fio_file *f)
1165 {
1166         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1167         assert(fio_file_open(f));
1168         f->references++;
1169 }
1170
1171 int put_file(struct thread_data *td, struct fio_file *f)
1172 {
1173         int f_ret = 0, ret = 0;
1174
1175         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1176
1177         if (!fio_file_open(f)) {
1178                 assert(f->fd == -1);
1179                 return 0;
1180         }
1181
1182         assert(f->references);
1183         if (--f->references)
1184                 return 0;
1185
1186         if (should_fsync(td) && td->o.fsync_on_close)
1187                 f_ret = fsync(f->fd);
1188
1189         if (td->io_ops->close_file)
1190                 ret = td->io_ops->close_file(td, f);
1191
1192         if (!ret)
1193                 ret = f_ret;
1194
1195         td->nr_open_files--;
1196         fio_file_clear_open(f);
1197         assert(f->fd == -1);
1198         return ret;
1199 }
1200
1201 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1202 {
1203         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1204                 return;
1205
1206         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1207                 if (ddir == DDIR_READ)
1208                         fio_rwlock_read(f->rwlock);
1209                 else
1210                         fio_rwlock_write(f->rwlock);
1211         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1212                 fio_mutex_down(f->lock);
1213
1214         td->file_locks[f->fileno] = td->o.file_lock_mode;
1215 }
1216
1217 void unlock_file(struct thread_data *td, struct fio_file *f)
1218 {
1219         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1220                 return;
1221
1222         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1223                 fio_rwlock_unlock(f->rwlock);
1224         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1225                 fio_mutex_up(f->lock);
1226
1227         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1228 }
1229
1230 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1231 {
1232         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1233                 unlock_file(td, f);
1234 }
1235
1236 static int recurse_dir(struct thread_data *td, const char *dirname)
1237 {
1238         struct dirent *dir;
1239         int ret = 0;
1240         DIR *D;
1241
1242         D = opendir(dirname);
1243         if (!D) {
1244                 char buf[FIO_VERROR_SIZE];
1245
1246                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1247                 td_verror(td, errno, buf);
1248                 return 1;
1249         }
1250
1251         while ((dir = readdir(D)) != NULL) {
1252                 char full_path[PATH_MAX];
1253                 struct stat sb;
1254
1255                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1256                         continue;
1257
1258                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1259
1260                 if (lstat(full_path, &sb) == -1) {
1261                         if (errno != ENOENT) {
1262                                 td_verror(td, errno, "stat");
1263                                 return 1;
1264                         }
1265                 }
1266
1267                 if (S_ISREG(sb.st_mode)) {
1268                         add_file(td, full_path);
1269                         td->o.nr_files++;
1270                         continue;
1271                 }
1272                 if (!S_ISDIR(sb.st_mode))
1273                         continue;
1274
1275                 ret = recurse_dir(td, full_path);
1276                 if (ret)
1277                         break;
1278         }
1279
1280         closedir(D);
1281         return ret;
1282 }
1283
1284 int add_dir_files(struct thread_data *td, const char *path)
1285 {
1286         int ret = recurse_dir(td, path);
1287
1288         if (!ret)
1289                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1290
1291         return ret;
1292 }
1293
1294 void dup_files(struct thread_data *td, struct thread_data *org)
1295 {
1296         struct fio_file *f;
1297         unsigned int i;
1298
1299         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1300
1301         if (!org->files)
1302                 return;
1303
1304         td->files = malloc(org->files_index * sizeof(f));
1305
1306         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1307                 td->file_locks = malloc(org->files_index);
1308
1309         for_each_file(org, f, i) {
1310                 struct fio_file *__f;
1311
1312                 __f = smalloc(sizeof(*__f));
1313                 if (!__f) {
1314                         log_err("fio: smalloc OOM\n");
1315                         assert(0);
1316                 }
1317                 __f->fd = -1;
1318                 fio_file_reset(td, __f);
1319
1320                 if (f->file_name) {
1321                         __f->file_name = smalloc_strdup(f->file_name);
1322                         if (!__f->file_name) {
1323                                 log_err("fio: smalloc OOM\n");
1324                                 assert(0);
1325                         }
1326
1327                         __f->filetype = f->filetype;
1328                 }
1329
1330                 td->files[i] = __f;
1331         }
1332 }
1333
1334 /*
1335  * Returns the index that matches the filename, or -1 if not there
1336  */
1337 int get_fileno(struct thread_data *td, const char *fname)
1338 {
1339         struct fio_file *f;
1340         unsigned int i;
1341
1342         for_each_file(td, f, i)
1343                 if (!strcmp(f->file_name, fname))
1344                         return i;
1345
1346         return -1;
1347 }
1348
1349 /*
1350  * For log usage, where we add/open/close files automatically
1351  */
1352 void free_release_files(struct thread_data *td)
1353 {
1354         close_files(td);
1355         td->files_index = 0;
1356         td->nr_normal_files = 0;
1357 }
1358
1359 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1360 {
1361         f->last_pos = f->file_offset;
1362         f->last_start = -1ULL;
1363         if (f->io_axmap)
1364                 axmap_reset(f->io_axmap);
1365         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1366                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1367 }