unix_io.c revision 544349270e4c74a6feb971123884a8cf5052a7ee
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 Public 14 * License. 15 * %End-Header% 16 */ 17 18#define _LARGEFILE_SOURCE 19#define _LARGEFILE64_SOURCE 20 21#include <stdio.h> 22#include <string.h> 23#if HAVE_UNISTD_H 24#include <unistd.h> 25#endif 26#if HAVE_ERRNO_H 27#include <errno.h> 28#endif 29#include <fcntl.h> 30#include <time.h> 31#ifdef __linux__ 32#include <sys/utsname.h> 33#endif 34#if HAVE_SYS_STAT_H 35#include <sys/stat.h> 36#endif 37#if HAVE_SYS_TYPES_H 38#include <sys/types.h> 39#endif 40#if HAVE_SYS_RESOURCE_H 41#include <sys/resource.h> 42#endif 43 44#include "ext2_fs.h" 45#include "ext2fs.h" 46 47/* 48 * For checking structure magic numbers... 49 */ 50 51#define EXT2_CHECK_MAGIC(struct, code) \ 52 if ((struct)->magic != (code)) return (code) 53 54struct unix_cache { 55 char *buf; 56 unsigned long block; 57 int access_time; 58 int dirty:1; 59 int in_use:1; 60}; 61 62#define CACHE_SIZE 8 63#define WRITE_DIRECT_SIZE 4 /* Must be smaller than CACHE_SIZE */ 64#define READ_DIRECT_SIZE 4 /* Should be smaller than CACHE_SIZE */ 65 66struct unix_private_data { 67 int magic; 68 int dev; 69 int flags; 70 int access_time; 71 struct unix_cache cache[CACHE_SIZE]; 72}; 73 74static errcode_t unix_open(const char *name, int flags, io_channel *channel); 75static errcode_t unix_close(io_channel channel); 76static errcode_t unix_set_blksize(io_channel channel, int blksize); 77static errcode_t unix_read_blk(io_channel channel, unsigned long block, 78 int count, void *data); 79static errcode_t unix_write_blk(io_channel channel, unsigned long block, 80 int count, const void *data); 81static errcode_t unix_flush(io_channel channel); 82static errcode_t unix_write_byte(io_channel channel, unsigned long offset, 83 int size, const void *data); 84 85static void reuse_cache(io_channel channel, struct unix_private_data *data, 86 struct unix_cache *cache, unsigned long block); 87 88static struct struct_io_manager struct_unix_manager = { 89 EXT2_ET_MAGIC_IO_MANAGER, 90 "Unix I/O Manager", 91 unix_open, 92 unix_close, 93 unix_set_blksize, 94 unix_read_blk, 95 unix_write_blk, 96 unix_flush, 97#ifdef __CYGWIN__ 98 0 99#else 100 unix_write_byte 101#endif 102}; 103 104io_manager unix_io_manager = &struct_unix_manager; 105 106/* 107 * Here are the raw I/O functions 108 */ 109#ifndef __CYGWIN__ 110static errcode_t raw_read_blk(io_channel channel, 111 struct unix_private_data *data, 112 unsigned long block, 113 int count, void *buf) 114{ 115 errcode_t retval; 116 ssize_t size; 117 ext2_loff_t location; 118 int actual = 0; 119 120 size = (count < 0) ? -count : count * channel->block_size; 121 location = (ext2_loff_t) block * channel->block_size; 122 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) { 123 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED; 124 goto error_out; 125 } 126 actual = read(data->dev, buf, size); 127 if (actual != size) { 128 if (actual < 0) 129 actual = 0; 130 retval = EXT2_ET_SHORT_READ; 131 goto error_out; 132 } 133 return 0; 134 135error_out: 136 memset((char *) buf+actual, 0, size-actual); 137 if (channel->read_error) 138 retval = (channel->read_error)(channel, block, count, buf, 139 size, actual, retval); 140 return retval; 141} 142#else /* __CYGWIN__ */ 143/* 144 * Windows block devices only allow sector alignment IO in offset and size 145 */ 146static errcode_t raw_read_blk(io_channel channel, 147 struct unix_private_data *data, 148 unsigned long block, 149 int count, void *buf) 150{ 151 errcode_t retval; 152 size_t size, alignsize, fragment; 153 ext2_loff_t location; 154 int total = 0, actual; 155#define BLOCKALIGN 512 156 char sector[BLOCKALIGN]; 157 158 size = (count < 0) ? -count : count * channel->block_size; 159 location = (ext2_loff_t) block * channel->block_size; 160#ifdef DEBUG 161 printf("count=%d, size=%d, block=%d, blk_size=%d, location=%lx\n", 162 count, size, block, channel->block_size, location); 163#endif 164 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) { 165 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED; 166 goto error_out; 167 } 168 fragment = size % BLOCKALIGN; 169 alignsize = size - fragment; 170 if (alignsize) { 171 actual = read(data->dev, buf, alignsize); 172 if (actual != alignsize) 173 goto short_read; 174 } 175 if (fragment) { 176 actual = read(data->dev, sector, BLOCKALIGN); 177 if (actual != BLOCKALIGN) 178 goto short_read; 179 memcpy(buf+alignsize, sector, fragment); 180 } 181 return 0; 182 183short_read: 184 if (actual>0) 185 total += actual; 186 retval = EXT2_ET_SHORT_READ; 187 188error_out: 189 memset((char *) buf+total, 0, size-actual); 190 if (channel->read_error) 191 retval = (channel->read_error)(channel, block, count, buf, 192 size, actual, retval); 193 return retval; 194} 195#endif 196 197static errcode_t raw_write_blk(io_channel channel, 198 struct unix_private_data *data, 199 unsigned long block, 200 int count, const void *buf) 201{ 202 ssize_t size; 203 ext2_loff_t location; 204 int actual = 0; 205 errcode_t retval; 206 207 if (count == 1) 208 size = channel->block_size; 209 else { 210 if (count < 0) 211 size = -count; 212 else 213 size = count * channel->block_size; 214 } 215 216 location = (ext2_loff_t) block * channel->block_size; 217 if (ext2fs_llseek(data->dev, location, SEEK_SET) != location) { 218 retval = errno ? errno : EXT2_ET_LLSEEK_FAILED; 219 goto error_out; 220 } 221 222 actual = write(data->dev, buf, size); 223 if (actual != size) { 224 retval = EXT2_ET_SHORT_WRITE; 225 goto error_out; 226 } 227 return 0; 228 229error_out: 230 if (channel->write_error) 231 retval = (channel->write_error)(channel, block, count, buf, 232 size, actual, retval); 233 return retval; 234} 235 236 237/* 238 * Here we implement the cache functions 239 */ 240 241/* Allocate the cache buffers */ 242static errcode_t alloc_cache(io_channel channel, 243 struct unix_private_data *data) 244{ 245 errcode_t retval; 246 struct unix_cache *cache; 247 int i; 248 249 data->access_time = 0; 250 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 251 cache->block = 0; 252 cache->access_time = 0; 253 cache->dirty = 0; 254 cache->in_use = 0; 255 if ((retval = ext2fs_get_mem(channel->block_size, 256 &cache->buf))) 257 return retval; 258 } 259 return 0; 260} 261 262/* Free the cache buffers */ 263static void free_cache(struct unix_private_data *data) 264{ 265 struct unix_cache *cache; 266 int i; 267 268 data->access_time = 0; 269 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 270 cache->block = 0; 271 cache->access_time = 0; 272 cache->dirty = 0; 273 cache->in_use = 0; 274 if (cache->buf) 275 ext2fs_free_mem(&cache->buf); 276 cache->buf = 0; 277 } 278} 279 280#ifndef NO_IO_CACHE 281/* 282 * Try to find a block in the cache. If the block is not found, and 283 * eldest is a non-zero pointer, then fill in eldest with the cache 284 * entry to that should be reused. 285 */ 286static struct unix_cache *find_cached_block(struct unix_private_data *data, 287 unsigned long block, 288 struct unix_cache **eldest) 289{ 290 struct unix_cache *cache, *unused_cache, *oldest_cache; 291 int i; 292 293 unused_cache = oldest_cache = 0; 294 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 295 if (!cache->in_use) { 296 if (!unused_cache) 297 unused_cache = cache; 298 continue; 299 } 300 if (cache->block == block) { 301 cache->access_time = ++data->access_time; 302 return cache; 303 } 304 if (!oldest_cache || 305 (cache->access_time < oldest_cache->access_time)) 306 oldest_cache = cache; 307 } 308 if (eldest) 309 *eldest = (unused_cache) ? unused_cache : oldest_cache; 310 return 0; 311} 312 313/* 314 * Reuse a particular cache entry for another block. 315 */ 316static void reuse_cache(io_channel channel, struct unix_private_data *data, 317 struct unix_cache *cache, unsigned long block) 318{ 319 if (cache->dirty && cache->in_use) 320 raw_write_blk(channel, data, cache->block, 1, cache->buf); 321 322 cache->in_use = 1; 323 cache->dirty = 0; 324 cache->block = block; 325 cache->access_time = ++data->access_time; 326} 327 328/* 329 * Flush all of the blocks in the cache 330 */ 331static errcode_t flush_cached_blocks(io_channel channel, 332 struct unix_private_data *data, 333 int invalidate) 334 335{ 336 struct unix_cache *cache; 337 errcode_t retval, retval2; 338 int i; 339 340 retval2 = 0; 341 for (i=0, cache = data->cache; i < CACHE_SIZE; i++, cache++) { 342 if (!cache->in_use) 343 continue; 344 345 if (invalidate) 346 cache->in_use = 0; 347 348 if (!cache->dirty) 349 continue; 350 351 retval = raw_write_blk(channel, data, 352 cache->block, 1, cache->buf); 353 if (retval) 354 retval2 = retval; 355 else 356 cache->dirty = 0; 357 } 358 return retval2; 359} 360#endif /* NO_IO_CACHE */ 361 362static errcode_t unix_open(const char *name, int flags, io_channel *channel) 363{ 364 io_channel io = NULL; 365 struct unix_private_data *data = NULL; 366 errcode_t retval; 367 int open_flags; 368 struct stat st; 369#ifdef __linux__ 370 struct utsname ut; 371#endif 372 373 if (name == 0) 374 return EXT2_ET_BAD_DEVICE_NAME; 375 retval = ext2fs_get_mem(sizeof(struct struct_io_channel), &io); 376 if (retval) 377 return retval; 378 memset(io, 0, sizeof(struct struct_io_channel)); 379 io->magic = EXT2_ET_MAGIC_IO_CHANNEL; 380 retval = ext2fs_get_mem(sizeof(struct unix_private_data), &data); 381 if (retval) 382 goto cleanup; 383 384 io->manager = unix_io_manager; 385 retval = ext2fs_get_mem(strlen(name)+1, &io->name); 386 if (retval) 387 goto cleanup; 388 389 strcpy(io->name, name); 390 io->private_data = data; 391 io->block_size = 1024; 392 io->read_error = 0; 393 io->write_error = 0; 394 io->refcount = 1; 395 396 memset(data, 0, sizeof(struct unix_private_data)); 397 data->magic = EXT2_ET_MAGIC_UNIX_IO_CHANNEL; 398 399 if ((retval = alloc_cache(io, data))) 400 goto cleanup; 401 402 open_flags = (flags & IO_FLAG_RW) ? O_RDWR : O_RDONLY; 403#ifdef HAVE_OPEN64 404 data->dev = open64(name, open_flags); 405#else 406 data->dev = open(name, open_flags); 407#endif 408 if (data->dev < 0) { 409 retval = errno; 410 goto cleanup; 411 } 412 413#ifdef __linux__ 414#undef RLIM_INFINITY 415#if (defined(__alpha__) || ((defined(__sparc__) || defined(__mips__)) && (SIZEOF_LONG == 4))) 416#define RLIM_INFINITY ((unsigned long)(~0UL>>1)) 417#else 418#define RLIM_INFINITY (~0UL) 419#endif 420 /* 421 * Work around a bug in 2.4.10-2.4.18 kernels where writes to 422 * block devices are wrongly getting hit by the filesize 423 * limit. This workaround isn't perfect, since it won't work 424 * if glibc wasn't built against 2.2 header files. (Sigh.) 425 * 426 */ 427 if ((flags & IO_FLAG_RW) && 428 (uname(&ut) == 0) && 429 ((ut.release[0] == '2') && (ut.release[1] == '.') && 430 (ut.release[2] == '4') && (ut.release[3] == '.') && 431 (ut.release[4] == '1') && (ut.release[5] >= '0') && 432 (ut.release[5] < '8')) && 433 (fstat(data->dev, &st) == 0) && 434 (S_ISBLK(st.st_mode))) { 435 struct rlimit rlim; 436 437 rlim.rlim_cur = rlim.rlim_max = (unsigned long) RLIM_INFINITY; 438 setrlimit(RLIMIT_FSIZE, &rlim); 439 getrlimit(RLIMIT_FSIZE, &rlim); 440 if (((unsigned long) rlim.rlim_cur) < 441 ((unsigned long) rlim.rlim_max)) { 442 rlim.rlim_cur = rlim.rlim_max; 443 setrlimit(RLIMIT_FSIZE, &rlim); 444 } 445 } 446#endif 447 *channel = io; 448 return 0; 449 450cleanup: 451 if (data) { 452 free_cache(data); 453 ext2fs_free_mem(&data); 454 } 455 if (io) 456 ext2fs_free_mem(&io); 457 return retval; 458} 459 460static errcode_t unix_close(io_channel channel) 461{ 462 struct unix_private_data *data; 463 errcode_t retval = 0; 464 465 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 466 data = (struct unix_private_data *) channel->private_data; 467 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 468 469 if (--channel->refcount > 0) 470 return 0; 471 472#ifndef NO_IO_CACHE 473 retval = flush_cached_blocks(channel, data, 0); 474#endif 475 476 if (close(data->dev) < 0) 477 retval = errno; 478 free_cache(data); 479 480 ext2fs_free_mem(&channel->private_data); 481 if (channel->name) 482 ext2fs_free_mem(&channel->name); 483 ext2fs_free_mem(&channel); 484 return retval; 485} 486 487static errcode_t unix_set_blksize(io_channel channel, int blksize) 488{ 489 struct unix_private_data *data; 490 errcode_t retval; 491 492 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 493 data = (struct unix_private_data *) channel->private_data; 494 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 495 496 if (channel->block_size != blksize) { 497#ifndef NO_IO_CACHE 498 if ((retval = flush_cached_blocks(channel, data, 0))) 499 return retval; 500#endif 501 502 channel->block_size = blksize; 503 free_cache(data); 504 if ((retval = alloc_cache(channel, data))) 505 return retval; 506 } 507 return 0; 508} 509 510 511static errcode_t unix_read_blk(io_channel channel, unsigned long block, 512 int count, void *buf) 513{ 514 struct unix_private_data *data; 515 struct unix_cache *cache, *reuse[READ_DIRECT_SIZE]; 516 errcode_t retval; 517 char *cp; 518 int i, j; 519 520 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 521 data = (struct unix_private_data *) channel->private_data; 522 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 523 524#ifdef NO_IO_CACHE 525 return raw_read_blk(channel, data, block, count, buf); 526#else 527 /* 528 * If we're doing an odd-sized read or a very large read, 529 * flush out the cache and then do a direct read. 530 */ 531 if (count < 0 || count > WRITE_DIRECT_SIZE) { 532 if ((retval = flush_cached_blocks(channel, data, 0))) 533 return retval; 534 return raw_read_blk(channel, data, block, count, buf); 535 } 536 537 cp = buf; 538 while (count > 0) { 539 /* If it's in the cache, use it! */ 540 if ((cache = find_cached_block(data, block, &reuse[0]))) { 541#ifdef DEBUG 542 printf("Using cached block %d\n", block); 543#endif 544 memcpy(cp, cache->buf, channel->block_size); 545 count--; 546 block++; 547 cp += channel->block_size; 548 continue; 549 } 550 /* 551 * Find the number of uncached blocks so we can do a 552 * single read request 553 */ 554 for (i=1; i < count; i++) 555 if (find_cached_block(data, block+i, &reuse[i])) 556 break; 557#ifdef DEBUG 558 printf("Reading %d blocks starting at %d\n", i, block); 559#endif 560 if ((retval = raw_read_blk(channel, data, block, i, cp))) 561 return retval; 562 563 /* Save the results in the cache */ 564 for (j=0; j < i; j++) { 565 count--; 566 cache = reuse[j]; 567 reuse_cache(channel, data, cache, block++); 568 memcpy(cache->buf, cp, channel->block_size); 569 cp += channel->block_size; 570 } 571 } 572 return 0; 573#endif /* NO_IO_CACHE */ 574} 575 576static errcode_t unix_write_blk(io_channel channel, unsigned long block, 577 int count, const void *buf) 578{ 579 struct unix_private_data *data; 580 struct unix_cache *cache, *reuse; 581 errcode_t retval = 0; 582 const char *cp; 583 int writethrough; 584 585 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 586 data = (struct unix_private_data *) channel->private_data; 587 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 588 589#ifdef NO_IO_CACHE 590 return raw_write_blk(channel, data, block, count, buf); 591#else 592 /* 593 * If we're doing an odd-sized write or a very large write, 594 * flush out the cache completely and then do a direct write. 595 */ 596 if (count < 0 || count > WRITE_DIRECT_SIZE) { 597 if ((retval = flush_cached_blocks(channel, data, 1))) 598 return retval; 599 return raw_write_blk(channel, data, block, count, buf); 600 } 601 602 /* 603 * For a moderate-sized multi-block write, first force a write 604 * if we're in write-through cache mode, and then fill the 605 * cache with the blocks. 606 */ 607 writethrough = channel->flags & CHANNEL_FLAGS_WRITETHROUGH; 608 if (writethrough) 609 retval = raw_write_blk(channel, data, block, count, buf); 610 611 cp = buf; 612 while (count > 0) { 613 cache = find_cached_block(data, block, &reuse); 614 if (!cache) { 615 cache = reuse; 616 reuse_cache(channel, data, cache, block); 617 } 618 memcpy(cache->buf, cp, channel->block_size); 619 cache->dirty = !writethrough; 620 count--; 621 block++; 622 cp += channel->block_size; 623 } 624 return retval; 625#endif /* NO_IO_CACHE */ 626} 627 628static errcode_t unix_write_byte(io_channel channel, unsigned long offset, 629 int size, const void *buf) 630{ 631 struct unix_private_data *data; 632 errcode_t retval = 0; 633 ssize_t actual; 634 635 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 636 data = (struct unix_private_data *) channel->private_data; 637 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 638 639#ifndef NO_IO_CACHE 640 /* 641 * Flush out the cache completely 642 */ 643 if ((retval = flush_cached_blocks(channel, data, 1))) 644 return retval; 645#endif 646 647 if (lseek(data->dev, offset, SEEK_SET) < 0) 648 return errno; 649 650 actual = write(data->dev, buf, size); 651 if (actual != size) 652 return EXT2_ET_SHORT_WRITE; 653 654 return 0; 655} 656 657/* 658 * Flush data buffers to disk. 659 */ 660static errcode_t unix_flush(io_channel channel) 661{ 662 struct unix_private_data *data; 663 errcode_t retval = 0; 664 665 EXT2_CHECK_MAGIC(channel, EXT2_ET_MAGIC_IO_CHANNEL); 666 data = (struct unix_private_data *) channel->private_data; 667 EXT2_CHECK_MAGIC(data, EXT2_ET_MAGIC_UNIX_IO_CHANNEL); 668 669#ifndef NO_IO_CACHE 670 retval = flush_cached_blocks(channel, data, 0); 671#endif 672 fsync(data->dev); 673 return retval; 674} 675 676