1 /* 2 * unix_io.c --- This is the Unix (well, really POSIX) implementation 3 * of the I/O manager. 4 * 5 * Implements a one-block write-through cache. 6 * 7 * Includes support for Windows NT support under Cygwin. 8 * 9 * Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 10 * 2002 by Theodore Ts'o. 11 * 12 * %Begin-Header% 13 * This file may be redistributed under the terms of the GNU Library 14 * General Public License, version 2. 15 * %End-Header% 16 */ 17 18 #if !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__) 19 #define _XOPEN_SOURCE 600 20 #define _DARWIN_C_SOURCE 21 #define _FILE_OFFSET_BITS 64 22 #ifndef _LARGEFILE_SOURCE 23 #define _LARGEFILE_SOURCE 24 #endif 25 #ifndef _LARGEFILE64_SOURCE 26 #define _LARGEFILE64_SOURCE 27 #endif 28 #ifndef _GNU_SOURCE 29 #define _GNU_SOURCE 30 #endif 31 #endif 32 33 #include "config.h" 34 #include <stdio.h> 35 #include <string.h> 36 #if HAVE_UNISTD_H 37 #include <unistd.h> 38 #endif 39 #if HAVE_ERRNO_H 40 #include <errno.h> 41 #endif 42 #include <fcntl.h> 43 #include <time.h> 44 #ifdef __linux__ 45 #include <sys/utsname.h> 46 #endif 47 #if HAVE_SYS_TYPES_H 48 #include <sys/types.h> 49 #endif 50 #ifdef HAVE_SYS_IOCTL_H 51 #include <sys/ioctl.h> 52 #endif 53 #ifdef HAVE_SYS_MOUNT_H 54 #include <sys/mount.h> 55 #endif 56 #if HAVE_SYS_STAT_H 57 #include <sys/stat.h> 58 #endif 59 #if HAVE_SYS_RESOURCE_H 60 #include <sys/resource.h> 61 #endif 62 #if HAVE_LINUX_FALLOC_H 63 #include <linux/falloc.h> 64 #endif 65 66 #if defined(__linux__) && defined(_IO) && !defined(BLKROGET) 67 #define BLKROGET _IO(0x12, 94) /* Get read-only status (0 = read_write). */ 68 #endif 69 70 #undef ALIGN_DEBUG 71 72 #include "ext2_fs.h" 73 #include "ext2fs.h" 74 75 /* 76 * For checking structure magic numbers... 77 */ 78 79 #define EXT2_CHECK_MAGIC(struct, code) \ 80 if ((struct)->magic != (code)) return (code) 81 82 struct unix_cache { 83 char *buf; 84 unsigned long long block; 85 int access_time; 86 unsigned dirty:1; 87 unsigned in_use:1; 88 }; 89 90 #define CACHE_SIZE 8 91 #define WRITE_DIRECT_SIZE 4 /* Must be smaller than CACHE_SIZE */ 92 #define READ_DIRECT_SIZE 4 /* Should be smaller than CACHE_SIZE */ 93 94 struct unix_private_data { 95 int magic; 96 int dev; 97 int flags; 98 int align; 99 int access_time; 100 ext2_loff_t offset; 101 struct unix_cache cache[CACHE_SIZE]; 102 void *bounce; 103 struct struct_io_stats io_stats; 104 }; 105 106 #define IS_ALIGNED(n, align) ((((uintptr_t) n) & \ 107 ((uintptr_t) ((align)-1))) == 0) 108 109 static errcode_t unix_get_stats(io_channel channel, io_stats *stats) 110 { 111 errcode_t retval = 0; 112 113 struct unix_private_data *data; 114 115 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 116 data = (struct unix_private_data *) channel->private_data; 117 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 118 119 if (stats) 120 *stats = &data->io_stats; 121 122 return retval; 123 } 124 125 /* 126 * Here are the raw I/O functions 127 */ 128 static errcode_t raw_read_blk(io_channel channel, 129 struct unix_private_data *data, 130 unsigned long long block, 131 int count, void *bufv) 132 { 133 errcode_t retval; 134 ssize_t size; 135 ext2_loff_t location; 136 int actual = 0; 137 unsigned char *buf = bufv; 138 139 size = (count < 0) ? -count : count * channel->block_size; 140 data->io_stats.bytes_read += size; 141 location = ((ext2_loff_t) block * channel->block_size) + data->offset; 142 143 #ifdef HAVE_PREAD64 144 /* Try an aligned pread */ 145 if ((channel->align == 0) || 146 (IS_ALIGNED(buf, channel->align) && 147 IS_ALIGNED(size, channel->align))) { 148 actual = pread64(data->dev, buf, size, location); 149 if (actual == size) 150 return 0; 151 } 152 #elif HAVE_PREAD 153 /* Try an aligned pread */ 154 if ((sizeof(off_t) >= sizeof(ext2_loff_t)) && 155 ((channel->align == 0) || 156 (IS_ALIGNED(buf, channel->align) && 157 IS_ALIGNED(size, channel->align)))) { 158 actual = pread(data->dev, buf, size, location); 159 if (actual == size) 160 return 0; 161 } 162 #endif /* HAVE_PREAD */ 163 164 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) { 165 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED; 166 goto error_out; 167 } 168 if ((channel->align == 0) || 169 (IS_ALIGNED(buf, channel->align) && 170 IS_ALIGNED(size, channel->align))) { 171 actual = read(data->dev, buf, size); 172 if (actual != size) { 173 short_read: 174 if (actual < 0) 175 actual = 0; 176 retval = EXT2_ET_SHORT_READ; 177 goto error_out; 178 } 179 return 0; 180 } 181 182 #ifdef ALIGN_DEBUG 183 printf("raw_read_blk: O_DIRECT fallback: %p %lu\n", buf, 184 (unsigned long) size); 185 #endif 186 187 /* 188 * The buffer or size which we're trying to read isn't aligned 189 * to the O_DIRECT rules, so we need to do this the hard way... 190 */ 191 while (size > 0) { 192 actual = read(data->dev, data->bounce, channel->block_size); 193 if (actual != channel->block_size) 194 goto short_read; 195 actual = size; 196 if (size > channel->block_size) 197 actual = channel->block_size; 198 memcpy(buf, data->bounce, actual); 199 size -= actual; 200 buf += actual; 201 } 202 return 0; 203 204 error_out: 205 memset((char *) buf+actual, 0, size-actual); 206 if (channel->read_error) 207 retval = (channel->read_error)(channel, block, count, buf, 208 size, actual, retval); 209 return retval; 210 } 211 212 static errcode_t raw_write_blk(io_channel channel, 213 struct unix_private_data *data, 214 unsigned long long block, 215 int count, const void *bufv) 216 { 217 ssize_t size; 218 ext2_loff_t location; 219 int actual = 0; 220 errcode_t retval; 221 const unsigned char *buf = bufv; 222 223 if (count == 1) 224 size = channel->block_size; 225 else { 226 if (count < 0) 227 size = -count; 228 else 229 size = count * channel->block_size; 230 } 231 data->io_stats.bytes_written += size; 232 233 location = ((ext2_loff_t) block * channel->block_size) + data->offset; 234 235 #ifdef HAVE_PWRITE64 236 /* Try an aligned pwrite */ 237 if ((channel->align == 0) || 238 (IS_ALIGNED(buf, channel->align) && 239 IS_ALIGNED(size, channel->align))) { 240 actual = pwrite64(data->dev, buf, size, location); 241 if (actual == size) 242 return 0; 243 } 244 #elif HAVE_PWRITE 245 /* Try an aligned pwrite */ 246 if ((sizeof(off_t) >= sizeof(ext2_loff_t)) && 247 ((channel->align == 0) || 248 (IS_ALIGNED(buf, channel->align) && 249 IS_ALIGNED(size, channel->align)))) { 250 actual = pwrite(data->dev, buf, size, location); 251 if (actual == size) 252 return 0; 253 } 254 #endif /* HAVE_PWRITE */ 255 256 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) { 257 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED; 258 goto error_out; 259 } 260 261 if ((channel->align == 0) || 262 (IS_ALIGNED(buf, channel->align) && 263 IS_ALIGNED(size, channel->align))) { 264 actual = write(data->dev, buf, size); 265 if (actual != size) { 266 short_write: 267 retval = EXT2_ET_SHORT_WRITE; 268 goto error_out; 269 } 270 return 0; 271 } 272 273 #ifdef ALIGN_DEBUG 274 printf("raw_write_blk: O_DIRECT fallback: %p %lu\n", buf, 275 (unsigned long) size); 276 #endif 277 /* 278 * The buffer or size which we're trying to write isn't aligned 279 * to the O_DIRECT rules, so we need to do this the hard way... 280 */ 281 while (size > 0) { 282 if (size < channel->block_size) { 283 actual = read(data->dev, data->bounce, 284 channel->block_size); 285 if (actual != channel->block_size) { 286 retval = EXT2_ET_SHORT_READ; 287 goto error_out; 288 } 289 } 290 actual = size; 291 if (size > channel->block_size) 292 actual = channel->block_size; 293 memcpy(data->bounce, buf, actual); 294 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) { 295 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED; 296 goto error_out; 297 } 298 actual = write(data->dev, data->bounce, channel->block_size); 299 if (actual != channel->block_size) 300 goto short_write; 301 size -= actual; 302 buf += actual; 303 } 304 return 0; 305 306 error_out: 307 if (channel->write_error) 308 retval = (channel->write_error)(channel, block, count, buf, 309 size, actual, retval); 310 return retval; 311 } 312 313 314 /* 315 * Here we implement the cache functions 316 */ 317 318 /* Allocate the cache buffers */ 319 static errcode_t alloc_cache(io_channel channel, 320 struct unix_private_data *data) 321 { 322 errcode_t retval; 323 struct unix_cache *cache; 324 int i; 325 326 data->access_time = 0; 327 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 328 cache->block = 0; 329 cache->access_time = 0; 330 cache->dirty = 0; 331 cache->in_use = 0; 332 if (cache->buf) 333 ext2fs_free_mem(&cache->buf); 334 retval = io_channel_alloc_buf(channel, 0, &cache->buf); 335 if (retval) 336 return retval; 337 } 338 if (channel->align) { 339 if (data->bounce) 340 ext2fs_free_mem(&data->bounce); 341 retval = io_channel_alloc_buf(channel, 0, &data->bounce); 342 } 343 return retval; 344 } 345 346 /* Free the cache buffers */ 347 static void free_cache(struct unix_private_data *data) 348 { 349 struct unix_cache *cache; 350 int i; 351 352 data->access_time = 0; 353 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 354 cache->block = 0; 355 cache->access_time = 0; 356 cache->dirty = 0; 357 cache->in_use = 0; 358 if (cache->buf) 359 ext2fs_free_mem(&cache->buf); 360 } 361 if (data->bounce) 362 ext2fs_free_mem(&data->bounce); 363 } 364 365 #ifndef NO_IO_CACHE 366 /* 367 * Try to find a block in the cache. If the block is not found, and 368 * eldest is a non-zero pointer, then fill in eldest with the cache 369 * entry to that should be reused. 370 */ 371 static struct unix_cache *find_cached_block(struct unix_private_data *data, 372 unsigned long long block, 373 struct unix_cache **eldest) 374 { 375 struct unix_cache *cache, *unused_cache, *oldest_cache; 376 int i; 377 378 unused_cache = oldest_cache = 0; 379 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 380 if (!cache->in_use) { 381 if (!unused_cache) 382 unused_cache = cache; 383 continue; 384 } 385 if (cache->block == block) { 386 cache->access_time = ++data->access_time; 387 return cache; 388 } 389 if (!oldest_cache || 390 (cache->access_time < oldest_cache->access_time)) 391 oldest_cache = cache; 392 } 393 if (eldest) 394 *eldest = (unused_cache) ? unused_cache : oldest_cache; 395 return 0; 396 } 397 398 /* 399 * Reuse a particular cache entry for another block. 400 */ 401 static void reuse_cache(io_channel channel, struct unix_private_data *data, 402 struct unix_cache *cache, unsigned long long block) 403 { 404 if (cache->dirty && cache->in_use) 405 raw_write_blk(channel, data, cache->block, 1, cache->buf); 406 407 cache->in_use = 1; 408 cache->dirty = 0; 409 cache->block = block; 410 cache->access_time = ++data->access_time; 411 } 412 413 /* 414 * Flush all of the blocks in the cache 415 */ 416 static errcode_t flush_cached_blocks(io_channel channel, 417 struct unix_private_data *data, 418 int invalidate) 419 420 { 421 struct unix_cache *cache; 422 errcode_t retval, retval2; 423 int i; 424 425 retval2 = 0; 426 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 427 if (!cache->in_use) 428 continue; 429 430 if (invalidate) 431 cache->in_use = 0; 432 433 if (!cache->dirty) 434 continue; 435 436 retval = raw_write_blk(channel, data, 437 cache->block, 1, cache->buf); 438 if (retval) 439 retval2 = retval; 440 else 441 cache->dirty = 0; 442 } 443 return retval2; 444 } 445 #endif /* NO_IO_CACHE */ 446 447 #ifdef __linux__ 448 #ifndef BLKDISCARDZEROES 449 #define BLKDISCARDZEROES _IO(0x12,124) 450 #endif 451 #endif 452 453 int ext2fs_open_file(const char *pathname, int flags, mode_t mode) 454 { 455 if (mode) 456 #if defined(HAVE_OPEN64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED) 457 return open64(pathname, flags, mode); 458 else 459 return open64(pathname, flags); 460 #else 461 return open(pathname, flags, mode); 462 else 463 return open(pathname, flags); 464 #endif 465 } 466 467 int ext2fs_stat(const char *path, ext2fs_struct_stat *buf) 468 { 469 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED) 470 return stat64(path, buf); 471 #else 472 return stat(path, buf); 473 #endif 474 } 475 476 int ext2fs_fstat(int fd, ext2fs_struct_stat *buf) 477 { 478 #if defined(HAVE_FSTAT64) && !defined(__OSX_AVAILABLE_BUT_DEPRECATED) 479 return fstat64(fd, buf); 480 #else 481 return fstat(fd, buf); 482 #endif 483 } 484 485 486 static errcode_t unix_open_channel(const char *name, int fd, 487 int flags, io_channel *channel, 488 io_manager io_mgr) 489 { 490 io_channel io = NULL; 491 struct unix_private_data *data = NULL; 492 errcode_t retval; 493 ext2fs_struct_stat st; 494 #ifdef __linux__ 495 struct utsname ut; 496 #endif 497 498 retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io); 499 if (retval) 500 goto cleanup; 501 memset(io, 0, sizeof(struct struct_io_channel)); 502 io->magic = EXT2_ET_MAGIC_IO_CHANNEL; 503 retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data); 504 if (retval) 505 goto cleanup; 506 507 io->manager = io_mgr; 508 retval = ext2fs_get_mem(strlen(name)+1, &io->name); 509 if (retval) 510 goto cleanup; 511 512 strcpy(io->name, name); 513 io->private_data = data; 514 io->block_size = 1024; 515 io->read_error = 0; 516 io->write_error = 0; 517 io->refcount = 1; 518 519 memset(data, 0, sizeof(struct unix_private_data)); 520 data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL; 521 data->io_stats.num_fields = 2; 522 data->flags = flags; 523 data->dev = fd; 524 525 #if defined(O_DIRECT) 526 if (flags & IO_FLAG_DIRECT_IO) 527 io->align = ext2fs_get_dio_alignment(data->dev); 528 #elif defined(F_NOCACHE) 529 if (flags & IO_FLAG_DIRECT_IO) 530 io->align = 4096; 531 #endif 532 533 /* 534 * If the device is really a block device, then set the 535 * appropriate flag, otherwise we can set DISCARD_ZEROES flag 536 * because we are going to use punch hole instead of discard 537 * and if it succeed, subsequent read from sparse area returns 538 * zero. 539 */ 540 if (ext2fs_fstat(data->dev, &st) == 0) { 541 if (S_ISBLK(st.st_mode)) 542 io->flags |= CHANNEL_FLAGS_BLOCK_DEVICE; 543 else 544 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES; 545 } 546 547 #ifdef BLKDISCARDZEROES 548 { 549 int zeroes = 0; 550 if (ioctl(data->dev, BLKDISCARDZEROES, &zeroes) == 0 && 551 zeroes) 552 io->flags |= CHANNEL_FLAGS_DISCARD_ZEROES; 553 } 554 #endif 555 556 #if defined(__CYGWIN__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__) 557 /* 558 * Some operating systems require that the buffers be aligned, 559 * regardless of O_DIRECT 560 */ 561 if (!io->align) 562 io->align = 512; 563 #endif 564 565 566 if ((retval = alloc_cache(io, data))) 567 goto cleanup; 568 569 #ifdef BLKROGET 570 if (flags & IO_FLAG_RW) { 571 int error; 572 int readonly = 0; 573 574 /* Is the block device actually writable? */ 575 error = ioctl(data->dev, BLKROGET, &readonly); 576 if (!error && readonly) { 577 retval = EPERM; 578 goto cleanup; 579 } 580 } 581 #endif 582 583 #ifdef __linux__ 584 #undef RLIM_INFINITY 585 #if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4))) 586 #define RLIM_INFINITY ((unsigned long)(~0UL>>1)) 587 #else 588 #define RLIM_INFINITY (~0UL) 589 #endif 590 /* 591 * Work around a bug in 2.4.10-2.4.18 kernels where writes to 592 * block devices are wrongly getting hit by the filesize 593 * limit. This workaround isn't perfect, since it won't work 594 * if glibc wasn't built against 2.2 header files. (Sigh.) 595 * 596 */ 597 if ((flags & IO_FLAG_RW) && 598 (uname(&ut) == 0) && 599 ((ut.release[0] == '2') && (ut.release[1] == '.') && 600 (ut.release[2] == '4') && (ut.release[3] == '.') && 601 (ut.release[4] == '1') && (ut.release[5] >= '0') && 602 (ut.release[5] < '8')) && 603 (ext2fs_fstat(data->dev, &st) == 0) && 604 (S_ISBLK(st.st_mode))) { 605 struct rlimit rlim; 606 607 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY; 608 setrlimit(RLIMIT_FSIZE, &rlim); 609 getrlimit(RLIMIT_FSIZE, &rlim); 610 if (((unsigned long) rlim.rlim_cur) < 611 ((unsigned long) rlim.rlim_max)) { 612 rlim.rlim_cur = rlim.rlim_max; 613 setrlimit(RLIMIT_FSIZE, &rlim); 614 } 615 } 616 #endif 617 *channel = io; 618 return 0; 619 620 cleanup: 621 if (data) { 622 if (data->dev >= 0) 623 close(data->dev); 624 free_cache(data); 625 ext2fs_free_mem(&data); 626 } 627 if (io) { 628 if (io->name) { 629 ext2fs_free_mem(&io->name); 630 } 631 ext2fs_free_mem(&io); 632 } 633 return retval; 634 } 635 636 static errcode_t unixfd_open(const char *str_fd, int flags, 637 io_channel *channel) 638 { 639 int fd; 640 int fd_flags; 641 642 fd = atoi(str_fd); 643 #if defined(HAVE_FCNTL) 644 fd_flags = fcntl(fd, F_GETFD); 645 if (fd_flags == -1) 646 return -EBADF; 647 648 flags = 0; 649 if (fd_flags & O_RDWR) 650 flags |= IO_FLAG_RW; 651 if (fd_flags & O_EXCL) 652 flags |= IO_FLAG_EXCLUSIVE; 653 #if defined(O_DIRECT) 654 if (fd_flags & O_DIRECT) 655 flags |= IO_FLAG_DIRECT_IO; 656 #endif 657 #endif /* HAVE_FCNTL */ 658 659 return unix_open_channel(str_fd, fd, flags, channel, unixfd_io_manager); 660 } 661 662 static errcode_t unix_open(const char *name, int flags, 663 io_channel *channel) 664 { 665 int fd = -1; 666 int open_flags; 667 668 if (name == 0) 669 return EXT2_ET_BAD_DEVICE_NAME; 670 671 open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY; 672 if (flags & IO_FLAG_EXCLUSIVE) 673 open_flags |= O_EXCL; 674 #if defined(O_DIRECT) 675 if (flags & IO_FLAG_DIRECT_IO) 676 open_flags |= O_DIRECT; 677 #endif 678 fd = ext2fs_open_file(name, open_flags, 0); 679 if (fd < 0) 680 return errno; 681 #if defined(F_NOCACHE) && !defined(IO_DIRECT) 682 if (flags & IO_FLAG_DIRECT_IO) { 683 if (fcntl(fd, F_NOCACHE, 1) < 0) 684 return errno; 685 } 686 #endif 687 return unix_open_channel(name, fd, flags, channel, unix_io_manager); 688 } 689 690 static errcode_t unix_close(io_channel channel) 691 { 692 struct unix_private_data *data; 693 errcode_t retval = 0; 694 695 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 696 data = (struct unix_private_data *) channel->private_data; 697 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 698 699 if (--channel->refcount > 0) 700 return 0; 701 702 #ifndef NO_IO_CACHE 703 retval = flush_cached_blocks(channel, data, 0); 704 #endif 705 706 if (close(data->dev) < 0) 707 retval = errno; 708 free_cache(data); 709 710 ext2fs_free_mem(&channel->private_data); 711 if (channel->name) 712 ext2fs_free_mem(&channel->name); 713 ext2fs_free_mem(&channel); 714 return retval; 715 } 716 717 static errcode_t unix_set_blksize(io_channel channel, int blksize) 718 { 719 struct unix_private_data *data; 720 errcode_t retval; 721 722 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 723 data = (struct unix_private_data *) channel->private_data; 724 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 725 726 if (channel->block_size != blksize) { 727 #ifndef NO_IO_CACHE 728 if ((retval = flush_cached_blocks(channel, data, 0))) 729 return retval; 730 #endif 731 732 channel->block_size = blksize; 733 free_cache(data); 734 if ((retval = alloc_cache(channel, data))) 735 return retval; 736 } 737 return 0; 738 } 739 740 static errcode_t unix_read_blk64(io_channel channel, unsigned long long block, 741 int count, void *buf) 742 { 743 struct unix_private_data *data; 744 struct unix_cache *cache, *reuse[READ_DIRECT_SIZE]; 745 errcode_t retval; 746 char *cp; 747 int i, j; 748 749 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 750 data = (struct unix_private_data *) channel->private_data; 751 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 752 753 #ifdef NO_IO_CACHE 754 return raw_read_blk(channel, data, block, count, buf); 755 #else 756 /* 757 * If we're doing an odd-sized read or a very large read, 758 * flush out the cache and then do a direct read. 759 */ 760 if (count < 0 || count > WRITE_DIRECT_SIZE) { 761 if ((retval = flush_cached_blocks(channel, data, 0))) 762 return retval; 763 return raw_read_blk(channel, data, block, count, buf); 764 } 765 766 cp = buf; 767 while (count > 0) { 768 /* If it's in the cache, use it! */ 769 if ((cache = find_cached_block(data, block, &reuse[0]))) { 770 #ifdef DEBUG 771 printf("Using cached block %lu\n", block); 772 #endif 773 memcpy(cp, cache->buf, channel->block_size); 774 count--; 775 block++; 776 cp += channel->block_size; 777 continue; 778 } 779 if (count == 1) { 780 /* 781 * Special case where we read directly into the 782 * cache buffer; important in the O_DIRECT case 783 */ 784 cache = reuse[0]; 785 reuse_cache(channel, data, cache, block); 786 if ((retval = raw_read_blk(channel, data, block, 1, 787 cache->buf))) { 788 cache->in_use = 0; 789 return retval; 790 } 791 memcpy(cp, cache->buf, channel->block_size); 792 return 0; 793 } 794 795 /* 796 * Find the number of uncached blocks so we can do a 797 * single read request 798 */ 799 for (i=1; i < count; i++) 800 if (find_cached_block(data, block+i, &reuse[i])) 801 break; 802 #ifdef DEBUG 803 printf("Reading %d blocks starting at %lu\n", i, block); 804 #endif 805 if ((retval = raw_read_blk(channel, data, block, i, cp))) 806 return retval; 807 808 /* Save the results in the cache */ 809 for (j=0; j < i; j++) { 810 count--; 811 cache = reuse[j]; 812 reuse_cache(channel, data, cache, block++); 813 memcpy(cache->buf, cp, channel->block_size); 814 cp += channel->block_size; 815 } 816 } 817 return 0; 818 #endif /* NO_IO_CACHE */ 819 } 820 821 static errcode_t unix_read_blk(io_channel channel, unsigned long block, 822 int count, void *buf) 823 { 824 return unix_read_blk64(channel, block, count, buf); 825 } 826 827 static errcode_t unix_write_blk64(io_channel channel, unsigned long long block, 828 int count, const void *buf) 829 { 830 struct unix_private_data *data; 831 struct unix_cache *cache, *reuse; 832 errcode_t retval = 0; 833 const char *cp; 834 int writethrough; 835 836 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 837 data = (struct unix_private_data *) channel->private_data; 838 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 839 840 #ifdef NO_IO_CACHE 841 return raw_write_blk(channel, data, block, count, buf); 842 #else 843 /* 844 * If we're doing an odd-sized write or a very large write, 845 * flush out the cache completely and then do a direct write. 846 */ 847 if (count < 0 || count > WRITE_DIRECT_SIZE) { 848 if ((retval = flush_cached_blocks(channel, data, 1))) 849 return retval; 850 return raw_write_blk(channel, data, block, count, buf); 851 } 852 853 /* 854 * For a moderate-sized multi-block write, first force a write 855 * if we're in write-through cache mode, and then fill the 856 * cache with the blocks. 857 */ 858 writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH; 859 if (writethrough) 860 retval = raw_write_blk(channel, data, block, count, buf); 861 862 cp = buf; 863 while (count > 0) { 864 cache = find_cached_block(data, block, &reuse); 865 if (!cache) { 866 cache = reuse; 867 reuse_cache(channel, data, cache, block); 868 } 869 if (cache->buf != cp) 870 memcpy(cache->buf, cp, channel->block_size); 871 cache->dirty = !writethrough; 872 count--; 873 block++; 874 cp += channel->block_size; 875 } 876 return retval; 877 #endif /* NO_IO_CACHE */ 878 } 879 880 static errcode_t unix_cache_readahead(io_channel channel, 881 unsigned long long block, 882 unsigned long long count) 883 { 884 #ifdef POSIX_FADV_WILLNEED 885 struct unix_private_data *data; 886 887 data = (struct unix_private_data *)channel->private_data; 888 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 889 return posix_fadvise(data->dev, 890 (ext2_loff_t)block * channel->block_size + data->offset, 891 (ext2_loff_t)count * channel->block_size, 892 POSIX_FADV_WILLNEED); 893 #else 894 return EXT2_ET_OP_NOT_SUPPORTED; 895 #endif 896 } 897 898 static errcode_t unix_write_blk(io_channel channel, unsigned long block, 899 int count, const void *buf) 900 { 901 return unix_write_blk64(channel, block, count, buf); 902 } 903 904 static errcode_t unix_write_byte(io_channel channel, unsigned long offset, 905 int size, const void *buf) 906 { 907 struct unix_private_data *data; 908 errcode_t retval = 0; 909 ssize_t actual; 910 911 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 912 data = (struct unix_private_data *) channel->private_data; 913 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 914 915 if (channel->align != 0) { 916 #ifdef ALIGN_DEBUG 917 printf("unix_write_byte: O_DIRECT fallback\n"); 918 #endif 919 return EXT2_ET_UNIMPLEMENTED; 920 } 921 922 #ifndef NO_IO_CACHE 923 /* 924 * Flush out the cache completely 925 */ 926 if ((retval = flush_cached_blocks(channel, data, 1))) 927 return retval; 928 #endif 929 930 if (lseek(data->dev, offset + data->offset, SEEK_SET) < 0) 931 return errno; 932 933 actual = write(data->dev, buf, size); 934 if (actual != size) 935 return EXT2_ET_SHORT_WRITE; 936 937 return 0; 938 } 939 940 /* 941 * Flush data buffers to disk. 942 */ 943 static errcode_t unix_flush(io_channel channel) 944 { 945 struct unix_private_data *data; 946 errcode_t retval = 0; 947 948 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 949 data = (struct unix_private_data *) channel->private_data; 950 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 951 952 #ifndef NO_IO_CACHE 953 retval = flush_cached_blocks(channel, data, 0); 954 #endif 955 #ifdef HAVE_FSYNC 956 fsync(data->dev); 957 #endif 958 return retval; 959 } 960 961 static errcode_t unix_set_option(io_channel channel, const char *option, 962 const char *arg) 963 { 964 struct unix_private_data *data; 965 unsigned long long tmp; 966 char *end; 967 968 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 969 data = (struct unix_private_data *) channel->private_data; 970 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 971 972 if (!strcmp(option, "offset")) { 973 if (!arg) 974 return EXT2_ET_INVALID_ARGUMENT; 975 976 tmp = strtoull(arg, &end, 0); 977 if (*end) 978 return EXT2_ET_INVALID_ARGUMENT; 979 data->offset = tmp; 980 if (data->offset < 0) 981 return EXT2_ET_INVALID_ARGUMENT; 982 return 0; 983 } 984 return EXT2_ET_INVALID_ARGUMENT; 985 } 986 987 #if defined(__linux__) && !defined(BLKDISCARD) 988 #define BLKDISCARD _IO(0x12,119) 989 #endif 990 991 static errcode_t unix_discard(io_channel channel, unsigned long long block, 992 unsigned long long count) 993 { 994 struct unix_private_data *data; 995 int ret; 996 997 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 998 data = (struct unix_private_data *) channel->private_data; 999 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 1000 1001 if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) { 1002 #ifdef BLKDISCARD 1003 __u64 range[2]; 1004 1005 range[0] = (__u64)(block) * channel->block_size + data->offset; 1006 range[1] = (__u64)(count) * channel->block_size; 1007 1008 ret = ioctl(data->dev, BLKDISCARD, &range); 1009 #else 1010 goto unimplemented; 1011 #endif 1012 } else { 1013 #if defined(HAVE_FALLOCATE) && defined(FALLOC_FL_PUNCH_HOLE) 1014 /* 1015 * If we are not on block device, try to use punch hole 1016 * to reclaim free space. 1017 */ 1018 ret = fallocate(data->dev, 1019 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, 1020 (off_t)(block) * channel->block_size + data->offset, 1021 (off_t)(count) * channel->block_size); 1022 #else 1023 goto unimplemented; 1024 #endif 1025 } 1026 if (ret < 0) { 1027 if (errno == EOPNOTSUPP) 1028 goto unimplemented; 1029 return errno; 1030 } 1031 return 0; 1032 unimplemented: 1033 return EXT2_ET_UNIMPLEMENTED; 1034 } 1035 1036 /* parameters might not be used if OS doesn't support zeroout */ 1037 #pragma GCC diagnostic push 1038 #pragma GCC diagnostic ignored "-Wunused-parameter" 1039 static errcode_t unix_zeroout(io_channel channel, unsigned long long block, 1040 unsigned long long count) 1041 { 1042 struct unix_private_data *data; 1043 int ret; 1044 1045 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 1046 data = (struct unix_private_data *) channel->private_data; 1047 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 1048 1049 if (getenv("UNIX_IO_NOZEROOUT")) 1050 goto unimplemented; 1051 1052 if (channel->flags & CHANNEL_FLAGS_BLOCK_DEVICE) { 1053 /* Not implemented until the BLKZEROOUT mess is fixed */ 1054 goto unimplemented; 1055 } else { 1056 /* Regular file, try to use truncate/punch/zero. */ 1057 struct stat statbuf; 1058 1059 if (count == 0) 1060 return 0; 1061 /* 1062 * If we're trying to zero a range past the end of the file, 1063 * extend the file size, then truncate everything. 1064 */ 1065 ret = fstat(data->dev, &statbuf); 1066 if (ret) 1067 goto err; 1068 if ((unsigned long long) statbuf.st_size < 1069 (block + count) * channel->block_size + data->offset) { 1070 ret = ftruncate(data->dev, 1071 (block + count) * channel->block_size + data->offset); 1072 if (ret) 1073 goto err; 1074 } 1075 #if defined(HAVE_FALLOCATE) && (defined(FALLOC_FL_ZERO_RANGE) || \ 1076 (defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE))) 1077 #if defined(FALLOC_FL_PUNCH_HOLE) && defined(FALLOC_FL_KEEP_SIZE) 1078 ret = fallocate(data->dev, 1079 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE, 1080 (off_t)(block) * channel->block_size + data->offset, 1081 (off_t)(count) * channel->block_size); 1082 if (ret == 0) 1083 goto err; 1084 #endif 1085 #ifdef FALLOC_FL_ZERO_RANGE 1086 ret = fallocate(data->dev, 1087 FALLOC_FL_ZERO_RANGE, 1088 (off_t)(block) * channel->block_size + data->offset, 1089 (off_t)(count) * channel->block_size); 1090 #endif 1091 #else 1092 goto unimplemented; 1093 #endif /* HAVE_FALLOCATE && (ZERO_RANGE || (PUNCH_HOLE && KEEP_SIZE)) */ 1094 } 1095 err: 1096 if (ret < 0) { 1097 if (errno == EOPNOTSUPP) 1098 goto unimplemented; 1099 return errno; 1100 } 1101 return 0; 1102 unimplemented: 1103 return EXT2_ET_UNIMPLEMENTED; 1104 } 1105 #pragma GCC diagnostic pop 1106 1107 static struct struct_io_manager struct_unix_manager = { 1108 .magic = EXT2_ET_MAGIC_IO_MANAGER, 1109 .name = "Unix I/O Manager", 1110 .open = unix_open, 1111 .close = unix_close, 1112 .set_blksize = unix_set_blksize, 1113 .read_blk = unix_read_blk, 1114 .write_blk = unix_write_blk, 1115 .flush = unix_flush, 1116 .write_byte = unix_write_byte, 1117 .set_option = unix_set_option, 1118 .get_stats = unix_get_stats, 1119 .read_blk64 = unix_read_blk64, 1120 .write_blk64 = unix_write_blk64, 1121 .discard = unix_discard, 1122 .cache_readahead = unix_cache_readahead, 1123 .zeroout = unix_zeroout, 1124 }; 1125 1126 io_manager unix_io_manager = &struct_unix_manager; 1127 1128 static struct struct_io_manager struct_unixfd_manager = { 1129 .magic = EXT2_ET_MAGIC_IO_MANAGER, 1130 .name = "Unix fd I/O Manager", 1131 .open = unixfd_open, 1132 .close = unix_close, 1133 .set_blksize = unix_set_blksize, 1134 .read_blk = unix_read_blk, 1135 .write_blk = unix_write_blk, 1136 .flush = unix_flush, 1137 .write_byte = unix_write_byte, 1138 .set_option = unix_set_option, 1139 .get_stats = unix_get_stats, 1140 .read_blk64 = unix_read_blk64, 1141 .write_blk64 = unix_write_blk64, 1142 .discard = unix_discard, 1143 .cache_readahead = unix_cache_readahead, 1144 .zeroout = unix_zeroout, 1145 }; 1146 1147 io_manager unixfd_io_manager = &struct_unixfd_manager; 1148
