filesetup.c

   1 #include <unistd.h>
   2 #include <fcntl.h>
   3 #include <string.h>
   4 #include <assert.h>
   5 #include <dirent.h>
   6 #include <libgen.h>
   7 #include <sys/stat.h>
   8 #include <sys/mman.h>
   9 #include <sys/types.h>
  10
  11 #include "fio.h"
  12 #include "smalloc.h"
  13 #include "filehash.h"
  14 #include "options.h"
  15 #include "os/os.h"
  16 #include "hash.h"
  17 #include "lib/axmap.h"
  18
  19 #ifdef CONFIG_LINUX_FALLOCATE
  20 #include <linux/falloc.h>
  21 #endif
  22
  23 static int root_warn;
  24
  25 static FLIST_HEAD(filename_list);
  26
  27 static inline void clear_error(struct thread_data *td)
  28 {
  29         td->error = 0;
  30         td->verror[0] = '\0';
  31 }
  32
  33 /*
  34  * Leaves f->fd open on success, caller must close
  35  */
  36 static int extend_file(struct thread_data *td, struct fio_file *f)
  37 {
  38         int r, new_layout = 0, unlink_file = 0, flags;
  39         unsigned long long left;
  40         unsigned int bs;
  41         char *b = NULL;
  42
  43         if (read_only) {
  44                 log_err("fio: refusing extend of file due to read-only\n");
  45                 return 0;
  46         }
  47
  48         /*
  49          * check if we need to lay the file out complete again. fio
  50          * does that for operations involving reads, or for writes
  51          * where overwrite is set
  52          */
  53         if (td_read(td) ||
  54            (td_write(td) && td->o.overwrite && !td->o.file_append) ||
  55             (td_write(td) && td->io_ops->flags & FIO_NOEXTEND))
  56                 new_layout = 1;
  57         if (td_write(td) && !td->o.overwrite && !td->o.file_append)
  58                 unlink_file = 1;
  59
  60         if (unlink_file || new_layout) {
  61                 dprint(FD_FILE, "layout unlink %s\n", f->file_name);
  62                 if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
  63                         td_verror(td, errno, "unlink");
  64                         return 1;
  65                 }
  66         }
  67
  68         flags = O_WRONLY | O_CREAT;
  69         if (new_layout)
  70                 flags |= O_TRUNC;
  71
  72         dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
  73         f->fd = open(f->file_name, flags, 0644);
  74         if (f->fd < 0) {
  75                 td_verror(td, errno, "open");
  76                 return 1;
  77         }
  78
  79 #ifdef CONFIG_POSIX_FALLOCATE
  80         if (!td->o.fill_device) {
  81                 switch (td->o.fallocate_mode) {
  82                 case FIO_FALLOCATE_NONE:
  83                         break;
  84                 case FIO_FALLOCATE_POSIX:
  85                         dprint(FD_FILE, "posix_fallocate file %s size %llu\n",
  86                                  f->file_name,
  87                                  (unsigned long long) f->real_file_size);
  88
  89                         r = posix_fallocate(f->fd, 0, f->real_file_size);
  90                         if (r > 0) {
  91                                 log_err("fio: posix_fallocate fails: %s\n",
  92                                                 strerror(r));
  93                         }
  94                         break;
  95 #ifdef CONFIG_LINUX_FALLOCATE
  96                 case FIO_FALLOCATE_KEEP_SIZE:
  97                         dprint(FD_FILE,
  98                                 "fallocate(FALLOC_FL_KEEP_SIZE) "
  99                                 "file %s size %llu\n", f->file_name,
 100                                 (unsigned long long) f->real_file_size);
 101
 102                         r = fallocate(f->fd, FALLOC_FL_KEEP_SIZE, 0,
 103                                         f->real_file_size);
 104                         if (r != 0)
 105                                 td_verror(td, errno, "fallocate");
 106
 107                         break;
 108 #endif /* CONFIG_LINUX_FALLOCATE */
 109                 default:
 110                         log_err("fio: unknown fallocate mode: %d\n",
 111                                 td->o.fallocate_mode);
 112                         assert(0);
 113                 }
 114         }
 115 #endif /* CONFIG_POSIX_FALLOCATE */
 116
 117         if (!new_layout)
 118                 goto done;
 119
 120         /*
 121          * The size will be -1ULL when fill_device is used, so don't truncate
 122          * or fallocate this file, just write it
 123          */
 124         if (!td->o.fill_device) {
 125                 dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
 126                                         (unsigned long long) f->real_file_size);
 127                 if (ftruncate(f->fd, f->real_file_size) == -1) {
 128                         if (errno != EFBIG) {
 129                                 td_verror(td, errno, "ftruncate");
 130                                 goto err;
 131                         }
 132                 }
 133         }
 134
 135         b = malloc(td->o.max_bs[DDIR_WRITE]);
 136
 137         left = f->real_file_size;
 138         while (left && !td->terminate) {
 139                 bs = td->o.max_bs[DDIR_WRITE];
 140                 if (bs > left)
 141                         bs = left;
 142
 143                 fill_io_buffer(td, b, bs, bs);
 144
 145                 r = write(f->fd, b, bs);
 146
 147                 if (r > 0) {
 148                         left -= r;
 149                         continue;
 150                 } else {
 151                         if (r < 0) {
 152                                 int __e = errno;
 153
 154                                 if (__e == ENOSPC) {
 155                                         if (td->o.fill_device)
 156                                                 break;
 157                                         log_info("fio: ENOSPC on laying out "
 158                                                  "file, stopping\n");
 159                                         break;
 160                                 }
 161                                 td_verror(td, errno, "write");
 162                         } else
 163                                 td_verror(td, EIO, "write");
 164
 165                         break;
 166                 }
 167         }
 168
 169         if (td->terminate) {
 170                 dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
 171                 unlink(f->file_name);
 172         } else if (td->o.create_fsync) {
 173                 if (fsync(f->fd) < 0) {
 174                         td_verror(td, errno, "fsync");
 175                         goto err;
 176                 }
 177         }
 178         if (td->o.fill_device && !td_write(td)) {
 179                 fio_file_clear_size_known(f);
 180                 if (td_io_get_file_size(td, f))
 181                         goto err;
 182                 if (f->io_size > f->real_file_size)
 183                         f->io_size = f->real_file_size;
 184         }
 185
 186         free(b);
 187 done:
 188         return 0;
 189 err:
 190         close(f->fd);
 191         f->fd = -1;
 192         if (b)
 193                 free(b);
 194         return 1;
 195 }
 196
 197 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 198 {
 199         int ret = 0, r, did_open = 0, old_runstate;
 200         unsigned long long left;
 201         unsigned int bs;
 202         char *b;
 203
 204         if (td->io_ops->flags & FIO_PIPEIO)
 205                 return 0;
 206
 207         if (!fio_file_open(f)) {
 208                 if (td->io_ops->open_file(td, f)) {
 209                         log_err("fio: cannot pre-read, failed to open file\n");
 210                         return 1;
 211                 }
 212                 did_open = 1;
 213         }
 214
 215         old_runstate = td_bump_runstate(td, TD_PRE_READING);
 216
 217         bs = td->o.max_bs[DDIR_READ];
 218         b = malloc(bs);
 219         memset(b, 0, bs);
 220
 221         if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
 222                 td_verror(td, errno, "lseek");
 223                 log_err("fio: failed to lseek pre-read file\n");
 224                 ret = 1;
 225                 goto error;
 226         }
 227
 228         left = f->io_size;
 229
 230         while (left && !td->terminate) {
 231                 if (bs > left)
 232                         bs = left;
 233
 234                 r = read(f->fd, b, bs);
 235
 236                 if (r == (int) bs) {
 237                         left -= bs;
 238                         continue;
 239                 } else {
 240                         td_verror(td, EIO, "pre_read");
 241                         break;
 242                 }
 243         }
 244
 245 error:
 246         td_restore_runstate(td, old_runstate);
 247
 248         if (did_open)
 249                 td->io_ops->close_file(td, f);
 250
 251         free(b);
 252         return ret;
 253 }
 254
 255 static unsigned long long get_rand_file_size(struct thread_data *td)
 256 {
 257         unsigned long long ret, sized;
 258         unsigned long r;
 259
 260         if (td->o.use_os_rand) {
 261                 r = os_random_long(&td->file_size_state);
 262                 sized = td->o.file_size_high - td->o.file_size_low;
 263                 ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
 264         } else {
 265                 r = __rand(&td->__file_size_state);
 266                 sized = td->o.file_size_high - td->o.file_size_low;
 267                 ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
 268         }
 269
 270         ret += td->o.file_size_low;
 271         ret -= (ret % td->o.rw_min_bs);
 272         return ret;
 273 }
 274
 275 static int file_size(struct thread_data *td, struct fio_file *f)
 276 {
 277         struct stat st;
 278
 279         if (stat(f->file_name, &st) == -1) {
 280                 td_verror(td, errno, "fstat");
 281                 return 1;
 282         }
 283
 284         f->real_file_size = st.st_size;
 285         return 0;
 286 }
 287
 288 static int bdev_size(struct thread_data *td, struct fio_file *f)
 289 {
 290         unsigned long long bytes = 0;
 291         int r;
 292
 293         if (td->io_ops->open_file(td, f)) {
 294                 log_err("fio: failed opening blockdev %s for size check\n",
 295                         f->file_name);
 296                 return 1;
 297         }
 298
 299         r = blockdev_size(f, &bytes);
 300         if (r) {
 301                 td_verror(td, r, "blockdev_size");
 302                 goto err;
 303         }
 304
 305         if (!bytes) {
 306                 log_err("%s: zero sized block device?\n", f->file_name);
 307                 goto err;
 308         }
 309
 310         f->real_file_size = bytes;
 311         td->io_ops->close_file(td, f);
 312         return 0;
 313 err:
 314         td->io_ops->close_file(td, f);
 315         return 1;
 316 }
 317
 318 static int char_size(struct thread_data *td, struct fio_file *f)
 319 {
 320 #ifdef FIO_HAVE_CHARDEV_SIZE
 321         unsigned long long bytes = 0;
 322         int r;
 323
 324         if (td->io_ops->open_file(td, f)) {
 325                 log_err("fio: failed opening blockdev %s for size check\n",
 326                         f->file_name);
 327                 return 1;
 328         }
 329
 330         r = chardev_size(f, &bytes);
 331         if (r) {
 332                 td_verror(td, r, "chardev_size");
 333                 goto err;
 334         }
 335
 336         if (!bytes) {
 337                 log_err("%s: zero sized char device?\n", f->file_name);
 338                 goto err;
 339         }
 340
 341         f->real_file_size = bytes;
 342         td->io_ops->close_file(td, f);
 343         return 0;
 344 err:
 345         td->io_ops->close_file(td, f);
 346         return 1;
 347 #else
 348         f->real_file_size = -1ULL;
 349         return 0;
 350 #endif
 351 }
 352
 353 static int get_file_size(struct thread_data *td, struct fio_file *f)
 354 {
 355         int ret = 0;
 356
 357         if (fio_file_size_known(f))
 358                 return 0;
 359
 360         if (f->filetype == FIO_TYPE_FILE)
 361                 ret = file_size(td, f);
 362         else if (f->filetype == FIO_TYPE_BD)
 363                 ret = bdev_size(td, f);
 364         else if (f->filetype == FIO_TYPE_CHAR)
 365                 ret = char_size(td, f);
 366         else
 367                 f->real_file_size = -1;
 368
 369         if (ret)
 370                 return ret;
 371
 372         if (f->file_offset > f->real_file_size) {
 373                 log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
 374                                         (unsigned long long) f->file_offset,
 375                                         (unsigned long long) f->real_file_size);
 376                 return 1;
 377         }
 378
 379         fio_file_set_size_known(f);
 380         return 0;
 381 }
 382
 383 static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 384                                    unsigned long long off,
 385                                    unsigned long long len)
 386 {
 387         int ret = 0;
 388
 389         if (len == -1ULL)
 390                 len = f->io_size;
 391         if (off == -1ULL)
 392                 off = f->file_offset;
 393
 394         if (len == -1ULL || off == -1ULL)
 395                 return 0;
 396
 397         dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
 398                                                                 len);
 399
 400         if (f->mmap_ptr) {
 401                 ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 402 #ifdef FIO_MADV_FREE
 403                 if (f->filetype == FIO_TYPE_BD)
 404                         (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
 405 #endif
 406         } else if (f->filetype == FIO_TYPE_FILE) {
 407                 ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
 408         } else if (f->filetype == FIO_TYPE_BD) {
 409                 ret = blockdev_invalidate_cache(f);
 410                 if (ret < 0 && errno == EACCES && geteuid()) {
 411                         if (!root_warn) {
 412                                 log_err("fio: only root may flush block "
 413                                         "devices. Cache flush bypassed!\n");
 414                                 root_warn = 1;
 415                         }
 416                         ret = 0;
 417                 }
 418         } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
 419                 ret = 0;
 420
 421         /*
 422          * Cache flushing isn't a fatal condition, and we know it will
 423          * happen on some platforms where we don't have the proper
 424          * function to flush eg block device caches. So just warn and
 425          * continue on our way.
 426          */
 427         if (ret) {
 428                 log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errno));
 429                 ret = 0;
 430         }
 431
 432         return 0;
 433
 434 }
 435
 436 int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 437 {
 438         if (!fio_file_open(f))
 439                 return 0;
 440
 441         return __file_invalidate_cache(td, f, -1ULL, -1ULL);
 442 }
 443
 444 int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 445 {
 446         int ret = 0;
 447
 448         dprint(FD_FILE, "fd close %s\n", f->file_name);
 449
 450         remove_file_hash(f);
 451
 452         if (close(f->fd) < 0)
 453                 ret = errno;
 454
 455         f->fd = -1;
 456
 457         if (f->shadow_fd != -1) {
 458                 close(f->shadow_fd);
 459                 f->shadow_fd = -1;
 460         }
 461
 462         f->engine_data = 0;
 463         return ret;
 464 }
 465
 466 int file_lookup_open(struct fio_file *f, int flags)
 467 {
 468         struct fio_file *__f;
 469         int from_hash;
 470
 471         __f = lookup_file_hash(f->file_name);
 472         if (__f) {
 473                 dprint(FD_FILE, "found file in hash %s\n", f->file_name);
 474                 /*
 475                  * racy, need the __f->lock locked
 476                  */
 477                 f->lock = __f->lock;
 478                 from_hash = 1;
 479         } else {
 480                 dprint(FD_FILE, "file not found in hash %s\n", f->file_name);
 481                 from_hash = 0;
 482         }
 483
 484         f->fd = open(f->file_name, flags, 0600);
 485         return from_hash;
 486 }
 487
 488 static int file_close_shadow_fds(struct thread_data *td)
 489 {
 490         struct fio_file *f;
 491         int num_closed = 0;
 492         unsigned int i;
 493
 494         for_each_file(td, f, i) {
 495                 if (f->shadow_fd == -1)
 496                         continue;
 497
 498                 close(f->shadow_fd);
 499                 f->shadow_fd = -1;
 500                 num_closed++;
 501         }
 502
 503         return num_closed;
 504 }
 505
 506 int generic_open_file(struct thread_data *td, struct fio_file *f)
 507 {
 508         int is_std = 0;
 509         int flags = 0;
 510         int from_hash = 0;
 511
 512         dprint(FD_FILE, "fd open %s\n", f->file_name);
 513
 514         if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
 515                 log_err("fio: trim only applies to block device\n");
 516                 return 1;
 517         }
 518
 519         if (!strcmp(f->file_name, "-")) {
 520                 if (td_rw(td)) {
 521                         log_err("fio: can't read/write to stdin/out\n");
 522                         return 1;
 523                 }
 524                 is_std = 1;
 525
 526                 /*
 527                  * move output logging to stderr, if we are writing to stdout
 528                  */
 529                 if (td_write(td))
 530                         f_out = stderr;
 531         }
 532
 533         if (td_trim(td))
 534                 goto skip_flags;
 535         if (td->o.odirect)
 536                 flags |= OS_O_DIRECT;
 537         if (td->o.oatomic) {
 538                 if (!FIO_O_ATOMIC) {
 539                         td_verror(td, EINVAL, "OS does not support atomic IO");
 540                         return 1;
 541                 }
 542                 flags |= OS_O_DIRECT | FIO_O_ATOMIC;
 543         }
 544         if (td->o.sync_io)
 545                 flags |= O_SYNC;
 546         if (td->o.create_on_open)
 547                 flags |= O_CREAT;
 548 skip_flags:
 549         if (f->filetype != FIO_TYPE_FILE)
 550                 flags |= FIO_O_NOATIME;
 551
 552 open_again:
 553         if (td_write(td)) {
 554                 if (!read_only)
 555                         flags |= O_RDWR;
 556
 557                 if (f->filetype == FIO_TYPE_FILE)
 558                         flags |= O_CREAT;
 559
 560                 if (is_std)
 561                         f->fd = dup(STDOUT_FILENO);
 562                 else
 563                         from_hash = file_lookup_open(f, flags);
 564         } else if (td_read(td)) {
 565                 if (f->filetype == FIO_TYPE_CHAR && !read_only)
 566                         flags |= O_RDWR;
 567                 else
 568                         flags |= O_RDONLY;
 569
 570                 if (is_std)
 571                         f->fd = dup(STDIN_FILENO);
 572                 else
 573                         from_hash = file_lookup_open(f, flags);
 574         } else { //td trim
 575                 flags |= O_RDWR;
 576                 from_hash = file_lookup_open(f, flags);
 577         }
 578
 579         if (f->fd == -1) {
 580                 char buf[FIO_VERROR_SIZE];
 581                 int __e = errno;
 582
 583                 if (__e == EPERM && (flags & FIO_O_NOATIME)) {
 584                         flags &= ~FIO_O_NOATIME;
 585                         goto open_again;
 586                 }
 587                 if (__e == EMFILE && file_close_shadow_fds(td))
 588                         goto open_again;
 589
 590                 snprintf(buf, sizeof(buf), "open(%s)", f->file_name);
 591
 592                 if (__e == EINVAL && (flags & OS_O_DIRECT)) {
 593                         log_err("fio: looks like your file system does not " \
 594                                 "support direct=1/buffered=0\n");
 595                 }
 596
 597                 td_verror(td, __e, buf);
 598         }
 599
 600         if (!from_hash && f->fd != -1) {
 601                 if (add_file_hash(f)) {
 602                         int fio_unused ret;
 603
 604                         /*
 605                          * Stash away descriptor for later close. This is to
 606                          * work-around a "feature" on Linux, where a close of
 607                          * an fd that has been opened for write will trigger
 608                          * udev to call blkid to check partitions, fs id, etc.
 609                          * That pollutes the device cache, which can slow down
 610                          * unbuffered accesses.
 611                          */
 612                         if (f->shadow_fd == -1)
 613                                 f->shadow_fd = f->fd;
 614                         else {
 615                                 /*
 616                                  * OK to ignore, we haven't done anything
 617                                  * with it
 618                                  */
 619                                 ret = generic_close_file(td, f);
 620                         }
 621                         goto open_again;
 622                 }
 623         }
 624
 625         return 0;
 626 }
 627
 628 int generic_get_file_size(struct thread_data *td, struct fio_file *f)
 629 {
 630         return get_file_size(td, f);
 631 }
 632
 633 /*
 634  * open/close all files, so that ->real_file_size gets set
 635  */
 636 static int get_file_sizes(struct thread_data *td)
 637 {
 638         struct fio_file *f;
 639         unsigned int i;
 640         int err = 0;
 641
 642         for_each_file(td, f, i) {
 643                 dprint(FD_FILE, "get file size for %p/%d/%p\n", f, i,
 644                                                                 f->file_name);
 645
 646                 if (td_io_get_file_size(td, f)) {
 647                         if (td->error != ENOENT) {
 648                                 log_err("%s\n", td->verror);
 649                                 err = 1;
 650                         }
 651                         clear_error(td);
 652                 }
 653
 654                 if (f->real_file_size == -1ULL && td->o.size)
 655                         f->real_file_size = td->o.size / td->o.nr_files;
 656         }
 657
 658         return err;
 659 }
 660
 661 struct fio_mount {
 662         struct flist_head list;
 663         const char *base;
 664         char __base[256];
 665         unsigned int key;
 666 };
 667
 668 /*
 669  * Get free number of bytes for each file on each unique mount.
 670  */
 671 static unsigned long long get_fs_free_counts(struct thread_data *td)
 672 {
 673         struct flist_head *n, *tmp;
 674         unsigned long long ret = 0;
 675         struct fio_mount *fm;
 676         FLIST_HEAD(list);
 677         struct fio_file *f;
 678         unsigned int i;
 679
 680         for_each_file(td, f, i) {
 681                 struct stat sb;
 682                 char buf[256];
 683
 684                 if (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_CHAR) {
 685                         if (f->real_file_size != -1ULL)
 686                                 ret += f->real_file_size;
 687                         continue;
 688                 } else if (f->filetype != FIO_TYPE_FILE)
 689                         continue;
 690
 691                 strcpy(buf, f->file_name);
 692
 693                 if (stat(buf, &sb) < 0) {
 694                         if (errno != ENOENT)
 695                                 break;
 696                         strcpy(buf, ".");
 697                         if (stat(buf, &sb) < 0)
 698                                 break;
 699                 }
 700
 701                 fm = NULL;
 702                 flist_for_each(n, &list) {
 703                         fm = flist_entry(n, struct fio_mount, list);
 704                         if (fm->key == sb.st_dev)
 705                                 break;
 706
 707                         fm = NULL;
 708                 }
 709
 710                 if (fm)
 711                         continue;
 712
 713                 fm = malloc(sizeof(*fm));
 714                 strcpy(fm->__base, buf);
 715                 fm->base = basename(fm->__base);
 716                 fm->key = sb.st_dev;
 717                 flist_add(&fm->list, &list);
 718         }
 719
 720         flist_for_each_safe(n, tmp, &list) {
 721                 unsigned long long sz;
 722
 723                 fm = flist_entry(n, struct fio_mount, list);
 724                 flist_del(&fm->list);
 725
 726                 sz = get_fs_size(fm->base);
 727                 if (sz && sz != -1ULL)
 728                         ret += sz;
 729
 730                 free(fm);
 731         }
 732
 733         return ret;
 734 }
 735
 736 uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
 737 {
 738         struct thread_options *o = &td->o;
 739
 740         if (o->file_append && f->filetype == FIO_TYPE_FILE)
 741                 return f->real_file_size;
 742
 743         return td->o.start_offset +
 744                 (td->thread_number - 1) * td->o.offset_increment;
 745 }
 746
 747 /*
 748  * Open the files and setup files sizes, creating files if necessary.
 749  */
 750 int setup_files(struct thread_data *td)
 751 {
 752         unsigned long long total_size, extend_size;
 753         struct thread_options *o = &td->o;
 754         struct fio_file *f;
 755         unsigned int i, nr_fs_extra = 0;
 756         int err = 0, need_extend;
 757         int old_state;
 758         const unsigned int bs = td_min_bs(td);
 759         uint64_t fs = 0;
 760
 761         dprint(FD_FILE, "setup files\n");
 762
 763         old_state = td_bump_runstate(td, TD_SETTING_UP);
 764
 765         if (o->read_iolog_file)
 766                 goto done;
 767
 768         /*
 769          * if ioengine defines a setup() method, it's responsible for
 770          * opening the files and setting f->real_file_size to indicate
 771          * the valid range for that file.
 772          */
 773         if (td->io_ops->setup)
 774                 err = td->io_ops->setup(td);
 775         else
 776                 err = get_file_sizes(td);
 777
 778         if (err)
 779                 goto err_out;
 780
 781         /*
 782          * check sizes. if the files/devices do not exist and the size
 783          * isn't passed to fio, abort.
 784          */
 785         total_size = 0;
 786         for_each_file(td, f, i) {
 787                 if (f->real_file_size == -1ULL)
 788                         total_size = -1ULL;
 789                 else
 790                         total_size += f->real_file_size;
 791         }
 792
 793         if (o->fill_device)
 794                 td->fill_device_size = get_fs_free_counts(td);
 795
 796         /*
 797          * device/file sizes are zero and no size given, punt
 798          */
 799         if ((!total_size || total_size == -1ULL) && !o->size &&
 800             !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
 801             !(o->nr_files && (o->file_size_low || o->file_size_high))) {
 802                 log_err("%s: you need to specify size=\n", o->name);
 803                 td_verror(td, EINVAL, "total_file_size");
 804                 goto err_out;
 805         }
 806
 807         /*
 808          * Calculate per-file size and potential extra size for the
 809          * first files, if needed.
 810          */
 811         if (!o->file_size_low && o->nr_files) {
 812                 uint64_t all_fs;
 813
 814                 fs = o->size / o->nr_files;
 815                 all_fs = fs * o->nr_files;
 816
 817                 if (all_fs < o->size)
 818                         nr_fs_extra = (o->size - all_fs) / bs;
 819         }
 820
 821         /*
 822          * now file sizes are known, so we can set ->io_size. if size= is
 823          * not given, ->io_size is just equal to ->real_file_size. if size
 824          * is given, ->io_size is size / nr_files.
 825          */
 826         extend_size = total_size = 0;
 827         need_extend = 0;
 828         for_each_file(td, f, i) {
 829                 f->file_offset = get_start_offset(td, f);
 830
 831                 if (!o->file_size_low) {
 832                         /*
 833                          * no file size range given, file size is equal to
 834                          * total size divided by number of files. If that is
 835                          * zero, set it to the real file size. If the size
 836                          * doesn't divide nicely with the min blocksize,
 837                          * make the first files bigger.
 838                          */
 839                         f->io_size = fs;
 840                         if (nr_fs_extra) {
 841                                 nr_fs_extra--;
 842                                 f->io_size += bs;
 843                         }
 844
 845                         if (!f->io_size)
 846                                 f->io_size = f->real_file_size - f->file_offset;
 847                 } else if (f->real_file_size < o->file_size_low ||
 848                            f->real_file_size > o->file_size_high) {
 849                         if (f->file_offset > o->file_size_low)
 850                                 goto err_offset;
 851                         /*
 852                          * file size given. if it's fixed, use that. if it's a
 853                          * range, generate a random size in-between.
 854                          */
 855                         if (o->file_size_low == o->file_size_high)
 856                                 f->io_size = o->file_size_low - f->file_offset;
 857                         else {
 858                                 f->io_size = get_rand_file_size(td)
 859                                                 - f->file_offset;
 860                         }
 861                 } else
 862                         f->io_size = f->real_file_size - f->file_offset;
 863
 864                 if (f->io_size == -1ULL)
 865                         total_size = -1ULL;
 866                 else {
 867                         if (o->size_percent)
 868                                 f->io_size = (f->io_size * o->size_percent) / 100;
 869                         total_size += f->io_size;
 870                 }
 871
 872                 if (f->filetype == FIO_TYPE_FILE &&
 873                     (f->io_size + f->file_offset) > f->real_file_size &&
 874                     !(td->io_ops->flags & FIO_DISKLESSIO)) {
 875                         if (!o->create_on_open) {
 876                                 need_extend++;
 877                                 extend_size += (f->io_size + f->file_offset);
 878                         } else
 879                                 f->real_file_size = f->io_size + f->file_offset;
 880                         fio_file_set_extend(f);
 881                 }
 882         }
 883
 884         if (!o->size || o->size > total_size)
 885                 o->size = total_size;
 886
 887         /*
 888          * See if we need to extend some files
 889          */
 890         if (need_extend) {
 891                 temp_stall_ts = 1;
 892                 if (output_format == FIO_OUTPUT_NORMAL)
 893                         log_info("%s: Laying out IO file(s) (%u file(s) /"
 894                                  " %lluMB)\n", o->name, need_extend,
 895                                         extend_size >> 20);
 896
 897                 for_each_file(td, f, i) {
 898                         unsigned long long old_len = -1ULL, extend_len = -1ULL;
 899
 900                         if (!fio_file_extend(f))
 901                                 continue;
 902
 903                         assert(f->filetype == FIO_TYPE_FILE);
 904                         fio_file_clear_extend(f);
 905                         if (!o->fill_device) {
 906                                 old_len = f->real_file_size;
 907                                 extend_len = f->io_size + f->file_offset -
 908                                                 old_len;
 909                         }
 910                         f->real_file_size = (f->io_size + f->file_offset);
 911                         err = extend_file(td, f);
 912                         if (err)
 913                                 break;
 914
 915                         err = __file_invalidate_cache(td, f, old_len,
 916                                                                 extend_len);
 917                         close(f->fd);
 918                         f->fd = -1;
 919                         if (err)
 920                                 break;
 921                 }
 922                 temp_stall_ts = 0;
 923         }
 924
 925         if (err)
 926                 goto err_out;
 927
 928         if (!o->zone_size)
 929                 o->zone_size = o->size;
 930
 931         /*
 932          * iolog already set the total io size, if we read back
 933          * stored entries.
 934          */
 935         if (!o->read_iolog_file)
 936                 td->total_io_size = o->size * o->loops;
 937
 938 done:
 939         if (o->create_only)
 940                 td->done = 1;
 941
 942         td_restore_runstate(td, old_state);
 943         return 0;
 944 err_offset:
 945         log_err("%s: you need to specify valid offset=\n", o->name);
 946 err_out:
 947         td_restore_runstate(td, old_state);
 948         return 1;
 949 }
 950
 951 int pre_read_files(struct thread_data *td)
 952 {
 953         struct fio_file *f;
 954         unsigned int i;
 955
 956         dprint(FD_FILE, "pre_read files\n");
 957
 958         for_each_file(td, f, i) {
 959                 pre_read_file(td, f);
 960         }
 961
 962         return 1;
 963 }
 964
 965 static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 966 {
 967         unsigned int range_size, seed;
 968         unsigned long nranges;
 969         uint64_t file_size;
 970
 971         range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
 972         file_size = min(f->real_file_size, f->io_size);
 973
 974         nranges = (file_size + range_size - 1) / range_size;
 975
 976         seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
 977         if (!td->o.rand_repeatable)
 978                 seed = td->rand_seeds[4];
 979
 980         if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
 981                 zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
 982         else
 983                 pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
 984
 985         return 1;
 986 }
 987
 988 static int init_rand_distribution(struct thread_data *td)
 989 {
 990         struct fio_file *f;
 991         unsigned int i;
 992         int state;
 993
 994         if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
 995                 return 0;
 996
 997         state = td_bump_runstate(td, TD_SETTING_UP);
 998
 999         for_each_file(td, f, i)
1000                 __init_rand_distribution(td, f);
1001
1002         td_restore_runstate(td, state);
1003
1004         return 1;
1005 }
1006
1007 int init_random_map(struct thread_data *td)
1008 {
1009         unsigned long long blocks;
1010         struct fio_file *f;
1011         unsigned int i;
1012
1013         if (init_rand_distribution(td))
1014                 return 0;
1015         if (!td_random(td))
1016                 return 0;
1017
1018         for_each_file(td, f, i) {
1019                 uint64_t file_size = min(f->real_file_size, f->io_size);
1020
1021                 blocks = file_size / (unsigned long long) td->o.rw_min_bs;
1022
1023                 if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
1024                         unsigned long seed;
1025
1026                         seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
1027
1028                         if (!lfsr_init(&f->lfsr, blocks, seed, 0))
1029                                 continue;
1030                 } else if (!td->o.norandommap) {
1031                         f->io_axmap = axmap_new(blocks);
1032                         if (f->io_axmap)
1033                                 continue;
1034                 } else if (td->o.norandommap)
1035                         continue;
1036
1037                 if (!td->o.softrandommap) {
1038                         log_err("fio: failed allocating random map. If running"
1039                                 " a large number of jobs, try the 'norandommap'"
1040                                 " option or set 'softrandommap'. Or give"
1041                                 " a larger --alloc-size to fio.\n");
1042                         return 1;
1043                 }
1044
1045                 log_info("fio: file %s failed allocating random map. Running "
1046                          "job without.\n", f->file_name);
1047         }
1048
1049         return 0;
1050 }
1051
1052 void close_files(struct thread_data *td)
1053 {
1054         struct fio_file *f;
1055         unsigned int i;
1056
1057         for_each_file(td, f, i) {
1058                 if (fio_file_open(f))
1059                         td_io_close_file(td, f);
1060         }
1061 }
1062
1063 void close_and_free_files(struct thread_data *td)
1064 {
1065         struct fio_file *f;
1066         unsigned int i;
1067
1068         dprint(FD_FILE, "close files\n");
1069
1070         for_each_file(td, f, i) {
1071                 if (fio_file_open(f))
1072                         td_io_close_file(td, f);
1073
1074                 remove_file_hash(f);
1075
1076                 if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
1077                         dprint(FD_FILE, "free unlink %s\n", f->file_name);
1078                         unlink(f->file_name);
1079                 }
1080
1081                 sfree(f->file_name);
1082                 f->file_name = NULL;
1083                 axmap_free(f->io_axmap);
1084                 f->io_axmap = NULL;
1085                 sfree(f);
1086         }
1087
1088         td->o.filename = NULL;
1089         free(td->files);
1090         free(td->file_locks);
1091         td->files_index = 0;
1092         td->files = NULL;
1093         td->file_locks = NULL;
1094         td->o.file_lock_mode = FILE_LOCK_NONE;
1095         td->o.nr_files = 0;
1096 }
1097
1098 static void get_file_type(struct fio_file *f)
1099 {
1100         struct stat sb;
1101
1102         if (!strcmp(f->file_name, "-"))
1103                 f->filetype = FIO_TYPE_PIPE;
1104         else
1105                 f->filetype = FIO_TYPE_FILE;
1106
1107         /* \\.\ is the device namespace in Windows, where every file is
1108          * a block device */
1109         if (strncmp(f->file_name, "\\\\.\\", 4) == 0)
1110                 f->filetype = FIO_TYPE_BD;
1111
1112         if (!stat(f->file_name, &sb)) {
1113                 if (S_ISBLK(sb.st_mode))
1114                         f->filetype = FIO_TYPE_BD;
1115                 else if (S_ISCHR(sb.st_mode))
1116                         f->filetype = FIO_TYPE_CHAR;
1117                 else if (S_ISFIFO(sb.st_mode))
1118                         f->filetype = FIO_TYPE_PIPE;
1119         }
1120 }
1121
1122 static int __is_already_allocated(const char *fname)
1123 {
1124         struct flist_head *entry;
1125         char *filename;
1126
1127         if (flist_empty(&filename_list))
1128                 return 0;
1129
1130         flist_for_each(entry, &filename_list) {
1131                 filename = flist_entry(entry, struct file_name, list)->filename;
1132
1133                 if (strcmp(filename, fname) == 0)
1134                         return 1;
1135         }
1136
1137         return 0;
1138 }
1139
1140 static int is_already_allocated(const char *fname)
1141 {
1142         int ret;
1143
1144         fio_file_hash_lock();
1145         ret = __is_already_allocated(fname);
1146         fio_file_hash_unlock();
1147         return ret;
1148 }
1149
1150 static void set_already_allocated(const char *fname)
1151 {
1152         struct file_name *fn;
1153
1154         fn = malloc(sizeof(struct file_name));
1155         fn->filename = strdup(fname);
1156
1157         fio_file_hash_lock();
1158         if (!__is_already_allocated(fname)) {
1159                 flist_add_tail(&fn->list, &filename_list);
1160                 fn = NULL;
1161         }
1162         fio_file_hash_unlock();
1163
1164         if (fn) {
1165                 free(fn->filename);
1166                 free(fn);
1167         }
1168 }
1169
1170
1171 static void free_already_allocated(void)
1172 {
1173         struct flist_head *entry, *tmp;
1174         struct file_name *fn;
1175
1176         if (flist_empty(&filename_list))
1177                 return;
1178
1179         fio_file_hash_lock();
1180         flist_for_each_safe(entry, tmp, &filename_list) {
1181                 fn = flist_entry(entry, struct file_name, list);
1182                 free(fn->filename);
1183                 flist_del(&fn->list);
1184                 free(fn);
1185         }
1186
1187         fio_file_hash_unlock();
1188 }
1189
1190 static struct fio_file *alloc_new_file(struct thread_data *td)
1191 {
1192         struct fio_file *f;
1193
1194         f = smalloc(sizeof(*f));
1195         if (!f) {
1196                 log_err("fio: smalloc OOM\n");
1197                 assert(0);
1198                 return NULL;
1199         }
1200
1201         f->fd = -1;
1202         f->shadow_fd = -1;
1203         fio_file_reset(td, f);
1204         return f;
1205 }
1206
1207 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
1208 {
1209         int cur_files = td->files_index;
1210         char file_name[PATH_MAX];
1211         struct fio_file *f;
1212         int len = 0;
1213
1214         dprint(FD_FILE, "add file %s\n", fname);
1215
1216         if (td->o.directory)
1217                 len = set_name_idx(file_name, td->o.directory, numjob);
1218
1219         sprintf(file_name + len, "%s", fname);
1220
1221         /* clean cloned siblings using existing files */
1222         if (numjob && is_already_allocated(file_name))
1223                 return 0;
1224
1225         f = alloc_new_file(td);
1226
1227         if (td->files_size <= td->files_index) {
1228                 unsigned int new_size = td->o.nr_files + 1;
1229
1230                 dprint(FD_FILE, "resize file array to %d files\n", new_size);
1231
1232                 td->files = realloc(td->files, new_size * sizeof(f));
1233                 if (td->files == NULL) {
1234                         log_err("fio: realloc OOM\n");
1235                         assert(0);
1236                 }
1237                 if (td->o.file_lock_mode != FILE_LOCK_NONE) {
1238                         td->file_locks = realloc(td->file_locks, new_size);
1239                         if (!td->file_locks) {
1240                                 log_err("fio: realloc OOM\n");
1241                                 assert(0);
1242                         }
1243                         td->file_locks[cur_files] = FILE_LOCK_NONE;
1244                 }
1245                 td->files_size = new_size;
1246         }
1247         td->files[cur_files] = f;
1248         f->fileno = cur_files;
1249
1250         /*
1251          * init function, io engine may not be loaded yet
1252          */
1253         if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
1254                 f->real_file_size = -1ULL;
1255
1256         f->file_name = smalloc_strdup(file_name);
1257         if (!f->file_name) {
1258                 log_err("fio: smalloc OOM\n");
1259                 assert(0);
1260         }
1261
1262         get_file_type(f);
1263
1264         switch (td->o.file_lock_mode) {
1265         case FILE_LOCK_NONE:
1266                 break;
1267         case FILE_LOCK_READWRITE:
1268                 f->rwlock = fio_rwlock_init();
1269                 break;
1270         case FILE_LOCK_EXCLUSIVE:
1271                 f->lock = fio_mutex_init(FIO_MUTEX_UNLOCKED);
1272                 break;
1273         default:
1274                 log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
1275                 assert(0);
1276         }
1277
1278         td->files_index++;
1279         if (f->filetype == FIO_TYPE_FILE)
1280                 td->nr_normal_files++;
1281
1282         set_already_allocated(file_name);
1283
1284         if (inc)
1285                 td->o.nr_files++;
1286
1287         dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
1288                                                         cur_files);
1289
1290         return cur_files;
1291 }
1292
1293 int add_file_exclusive(struct thread_data *td, const char *fname)
1294 {
1295         struct fio_file *f;
1296         unsigned int i;
1297
1298         for_each_file(td, f, i) {
1299                 if (!strcmp(f->file_name, fname))
1300                         return i;
1301         }
1302
1303         return add_file(td, fname, 0, 1);
1304 }
1305
1306 void get_file(struct fio_file *f)
1307 {
1308         dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
1309         assert(fio_file_open(f));
1310         f->references++;
1311 }
1312
1313 int put_file(struct thread_data *td, struct fio_file *f)
1314 {
1315         int f_ret = 0, ret = 0;
1316
1317         dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
1318
1319         if (!fio_file_open(f)) {
1320                 assert(f->fd == -1);
1321                 return 0;
1322         }
1323
1324         assert(f->references);
1325         if (--f->references)
1326                 return 0;
1327
1328         if (should_fsync(td) && td->o.fsync_on_close)
1329                 f_ret = fsync(f->fd);
1330
1331         if (td->io_ops->close_file)
1332                 ret = td->io_ops->close_file(td, f);
1333
1334         if (!ret)
1335                 ret = f_ret;
1336
1337         td->nr_open_files--;
1338         fio_file_clear_open(f);
1339         assert(f->fd == -1);
1340         return ret;
1341 }
1342
1343 void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
1344 {
1345         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1346                 return;
1347
1348         if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
1349                 if (ddir == DDIR_READ)
1350                         fio_rwlock_read(f->rwlock);
1351                 else
1352                         fio_rwlock_write(f->rwlock);
1353         } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1354                 fio_mutex_down(f->lock);
1355
1356         td->file_locks[f->fileno] = td->o.file_lock_mode;
1357 }
1358
1359 void unlock_file(struct thread_data *td, struct fio_file *f)
1360 {
1361         if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
1362                 return;
1363
1364         if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1365                 fio_rwlock_unlock(f->rwlock);
1366         else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1367                 fio_mutex_up(f->lock);
1368
1369         td->file_locks[f->fileno] = FILE_LOCK_NONE;
1370 }
1371
1372 void unlock_file_all(struct thread_data *td, struct fio_file *f)
1373 {
1374         if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
1375                 return;
1376         if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
1377                 unlock_file(td, f);
1378 }
1379
1380 static int recurse_dir(struct thread_data *td, const char *dirname)
1381 {
1382         struct dirent *dir;
1383         int ret = 0;
1384         DIR *D;
1385
1386         D = opendir(dirname);
1387         if (!D) {
1388                 char buf[FIO_VERROR_SIZE];
1389
1390                 snprintf(buf, FIO_VERROR_SIZE, "opendir(%s)", dirname);
1391                 td_verror(td, errno, buf);
1392                 return 1;
1393         }
1394
1395         while ((dir = readdir(D)) != NULL) {
1396                 char full_path[PATH_MAX];
1397                 struct stat sb;
1398
1399                 if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
1400                         continue;
1401
1402                 sprintf(full_path, "%s%s%s", dirname, FIO_OS_PATH_SEPARATOR, dir->d_name);
1403
1404                 if (lstat(full_path, &sb) == -1) {
1405                         if (errno != ENOENT) {
1406                                 td_verror(td, errno, "stat");
1407                                 ret = 1;
1408                                 break;
1409                         }
1410                 }
1411
1412                 if (S_ISREG(sb.st_mode)) {
1413                         add_file(td, full_path, 0, 1);
1414                         continue;
1415                 }
1416                 if (!S_ISDIR(sb.st_mode))
1417                         continue;
1418
1419                 ret = recurse_dir(td, full_path);
1420                 if (ret)
1421                         break;
1422         }
1423
1424         closedir(D);
1425         return ret;
1426 }
1427
1428 int add_dir_files(struct thread_data *td, const char *path)
1429 {
1430         int ret = recurse_dir(td, path);
1431
1432         if (!ret)
1433                 log_info("fio: opendir added %d files\n", td->o.nr_files);
1434
1435         return ret;
1436 }
1437
1438 void dup_files(struct thread_data *td, struct thread_data *org)
1439 {
1440         struct fio_file *f;
1441         unsigned int i;
1442
1443         dprint(FD_FILE, "dup files: %d\n", org->files_index);
1444
1445         if (!org->files)
1446                 return;
1447
1448         td->files = malloc(org->files_index * sizeof(f));
1449
1450         if (td->o.file_lock_mode != FILE_LOCK_NONE)
1451                 td->file_locks = malloc(org->files_index);
1452
1453         for_each_file(org, f, i) {
1454                 struct fio_file *__f;
1455
1456                 __f = alloc_new_file(td);
1457
1458                 if (f->file_name) {
1459                         __f->file_name = smalloc_strdup(f->file_name);
1460                         if (!__f->file_name) {
1461                                 log_err("fio: smalloc OOM\n");
1462                                 assert(0);
1463                         }
1464
1465                         __f->filetype = f->filetype;
1466                 }
1467
1468                 if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
1469                         __f->lock = f->lock;
1470                 else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
1471                         __f->rwlock = f->rwlock;
1472
1473                 td->files[i] = __f;
1474         }
1475 }
1476
1477 /*
1478  * Returns the index that matches the filename, or -1 if not there
1479  */
1480 int get_fileno(struct thread_data *td, const char *fname)
1481 {
1482         struct fio_file *f;
1483         unsigned int i;
1484
1485         for_each_file(td, f, i)
1486                 if (!strcmp(f->file_name, fname))
1487                         return i;
1488
1489         return -1;
1490 }
1491
1492 /*
1493  * For log usage, where we add/open/close files automatically
1494  */
1495 void free_release_files(struct thread_data *td)
1496 {
1497         close_files(td);
1498         td->files_index = 0;
1499         td->nr_normal_files = 0;
1500 }
1501
1502 void fio_file_reset(struct thread_data *td, struct fio_file *f)
1503 {
1504         f->last_pos = f->file_offset;
1505         f->last_start = -1ULL;
1506         if (f->io_axmap)
1507                 axmap_reset(f->io_axmap);
1508         if (td->o.random_generator == FIO_RAND_GEN_LFSR)
1509                 lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
1510 }
1511
1512 int fio_files_done(struct thread_data *td)
1513 {
1514         struct fio_file *f;
1515         unsigned int i;
1516
1517         for_each_file(td, f, i)
1518                 if (!fio_file_done(f))
1519                         return 0;
1520
1521         return 1;
1522 }
1523
1524 /* free memory used in initialization phase only */
1525 void filesetup_mem_free(void)
1526 {
1527         free_already_allocated();
1528 }