Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 1 | /* -*- mode: c; c-basic-offset: 8; -*- |
| 2 | * vim: noexpandtab sw=8 ts=8 sts=0: |
| 3 | * |
| 4 | * blockcheck.c |
| 5 | * |
| 6 | * Checksum and ECC codes for the OCFS2 userspace library. |
| 7 | * |
| 8 | * Copyright (C) 2006, 2008 Oracle. All rights reserved. |
| 9 | * |
| 10 | * This program is free software; you can redistribute it and/or |
| 11 | * modify it under the terms of the GNU General Public |
| 12 | * License, version 2, as published by the Free Software Foundation. |
| 13 | * |
| 14 | * This program is distributed in the hope that it will be useful, |
| 15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 17 | * General Public License for more details. |
| 18 | */ |
| 19 | |
| 20 | #include <linux/kernel.h> |
| 21 | #include <linux/types.h> |
| 22 | #include <linux/crc32.h> |
| 23 | #include <linux/buffer_head.h> |
| 24 | #include <linux/bitops.h> |
| 25 | #include <asm/byteorder.h> |
| 26 | |
Joel Becker | d6b32bb | 2008-10-17 14:55:01 -0700 | [diff] [blame] | 27 | #include <cluster/masklog.h> |
| 28 | |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 29 | #include "ocfs2.h" |
| 30 | |
| 31 | #include "blockcheck.h" |
| 32 | |
| 33 | |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 34 | /* |
| 35 | * We use the following conventions: |
| 36 | * |
| 37 | * d = # data bits |
| 38 | * p = # parity bits |
| 39 | * c = # total code bits (d + p) |
| 40 | */ |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 41 | |
| 42 | /* |
| 43 | * Calculate the bit offset in the hamming code buffer based on the bit's |
| 44 | * offset in the data buffer. Since the hamming code reserves all |
| 45 | * power-of-two bits for parity, the data bit number and the code bit |
| 46 | * number are offest by all the parity bits beforehand. |
| 47 | * |
| 48 | * Recall that bit numbers in hamming code are 1-based. This function |
| 49 | * takes the 0-based data bit from the caller. |
| 50 | * |
| 51 | * An example. Take bit 1 of the data buffer. 1 is a power of two (2^0), |
| 52 | * so it's a parity bit. 2 is a power of two (2^1), so it's a parity bit. |
| 53 | * 3 is not a power of two. So bit 1 of the data buffer ends up as bit 3 |
| 54 | * in the code buffer. |
| 55 | */ |
| 56 | static unsigned int calc_code_bit(unsigned int i) |
| 57 | { |
| 58 | unsigned int b, p; |
| 59 | |
| 60 | /* |
| 61 | * Data bits are 0-based, but we're talking code bits, which |
| 62 | * are 1-based. |
| 63 | */ |
| 64 | b = i + 1; |
| 65 | |
| 66 | /* |
| 67 | * For every power of two below our bit number, bump our bit. |
| 68 | * |
| 69 | * We compare with (b + 1) becuase we have to compare with what b |
| 70 | * would be _if_ it were bumped up by the parity bit. Capice? |
| 71 | */ |
| 72 | for (p = 0; (1 << p) < (b + 1); p++) |
| 73 | b++; |
| 74 | |
| 75 | return b; |
| 76 | } |
| 77 | |
| 78 | /* |
| 79 | * This is the low level encoder function. It can be called across |
| 80 | * multiple hunks just like the crc32 code. 'd' is the number of bits |
| 81 | * _in_this_hunk_. nr is the bit offset of this hunk. So, if you had |
| 82 | * two 512B buffers, you would do it like so: |
| 83 | * |
| 84 | * parity = ocfs2_hamming_encode(0, buf1, 512 * 8, 0); |
| 85 | * parity = ocfs2_hamming_encode(parity, buf2, 512 * 8, 512 * 8); |
| 86 | * |
| 87 | * If you just have one buffer, use ocfs2_hamming_encode_block(). |
| 88 | */ |
| 89 | u32 ocfs2_hamming_encode(u32 parity, void *data, unsigned int d, unsigned int nr) |
| 90 | { |
Joel Becker | e798b3f | 2008-12-15 17:13:48 -0800 | [diff] [blame^] | 91 | unsigned int i, b; |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 92 | |
Joel Becker | e798b3f | 2008-12-15 17:13:48 -0800 | [diff] [blame^] | 93 | BUG_ON(!d); |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 94 | |
| 95 | /* |
| 96 | * b is the hamming code bit number. Hamming code specifies a |
| 97 | * 1-based array, but C uses 0-based. So 'i' is for C, and 'b' is |
| 98 | * for the algorithm. |
| 99 | * |
| 100 | * The i++ in the for loop is so that the start offset passed |
| 101 | * to ocfs2_find_next_bit_set() is one greater than the previously |
| 102 | * found bit. |
| 103 | */ |
| 104 | for (i = 0; (i = ocfs2_find_next_bit(data, d, i)) < d; i++) |
| 105 | { |
| 106 | /* |
| 107 | * i is the offset in this hunk, nr + i is the total bit |
| 108 | * offset. |
| 109 | */ |
| 110 | b = calc_code_bit(nr + i); |
| 111 | |
Joel Becker | e798b3f | 2008-12-15 17:13:48 -0800 | [diff] [blame^] | 112 | /* |
| 113 | * Data bits in the resultant code are checked by |
| 114 | * parity bits that are part of the bit number |
| 115 | * representation. Huh? |
| 116 | * |
| 117 | * <wikipedia href="http://en.wikipedia.org/wiki/Hamming_code"> |
| 118 | * In other words, the parity bit at position 2^k |
| 119 | * checks bits in positions having bit k set in |
| 120 | * their binary representation. Conversely, for |
| 121 | * instance, bit 13, i.e. 1101(2), is checked by |
| 122 | * bits 1000(2) = 8, 0100(2)=4 and 0001(2) = 1. |
| 123 | * </wikipedia> |
| 124 | * |
| 125 | * Note that 'k' is the _code_ bit number. 'b' in |
| 126 | * our loop. |
| 127 | */ |
| 128 | parity ^= b; |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 129 | } |
| 130 | |
| 131 | /* While the data buffer was treated as little endian, the |
| 132 | * return value is in host endian. */ |
| 133 | return parity; |
| 134 | } |
| 135 | |
| 136 | u32 ocfs2_hamming_encode_block(void *data, unsigned int blocksize) |
| 137 | { |
| 138 | return ocfs2_hamming_encode(0, data, blocksize * 8, 0); |
| 139 | } |
| 140 | |
| 141 | /* |
| 142 | * Like ocfs2_hamming_encode(), this can handle hunks. nr is the bit |
| 143 | * offset of the current hunk. If bit to be fixed is not part of the |
| 144 | * current hunk, this does nothing. |
| 145 | * |
| 146 | * If you only have one hunk, use ocfs2_hamming_fix_block(). |
| 147 | */ |
| 148 | void ocfs2_hamming_fix(void *data, unsigned int d, unsigned int nr, |
| 149 | unsigned int fix) |
| 150 | { |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 151 | unsigned int i, b; |
| 152 | |
Joel Becker | e798b3f | 2008-12-15 17:13:48 -0800 | [diff] [blame^] | 153 | BUG_ON(!d); |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 154 | |
| 155 | /* |
| 156 | * If the bit to fix has an hweight of 1, it's a parity bit. One |
| 157 | * busted parity bit is its own error. Nothing to do here. |
| 158 | */ |
| 159 | if (hweight32(fix) == 1) |
| 160 | return; |
| 161 | |
| 162 | /* |
| 163 | * nr + d is the bit right past the data hunk we're looking at. |
| 164 | * If fix after that, nothing to do |
| 165 | */ |
| 166 | if (fix >= calc_code_bit(nr + d)) |
| 167 | return; |
| 168 | |
| 169 | /* |
| 170 | * nr is the offset in the data hunk we're starting at. Let's |
| 171 | * start b at the offset in the code buffer. See hamming_encode() |
| 172 | * for a more detailed description of 'b'. |
| 173 | */ |
| 174 | b = calc_code_bit(nr); |
| 175 | /* If the fix is before this hunk, nothing to do */ |
| 176 | if (fix < b) |
| 177 | return; |
| 178 | |
| 179 | for (i = 0; i < d; i++, b++) |
| 180 | { |
| 181 | /* Skip past parity bits */ |
| 182 | while (hweight32(b) == 1) |
| 183 | b++; |
| 184 | |
| 185 | /* |
| 186 | * i is the offset in this data hunk. |
| 187 | * nr + i is the offset in the total data buffer. |
| 188 | * b is the offset in the total code buffer. |
| 189 | * |
| 190 | * Thus, when b == fix, bit i in the current hunk needs |
| 191 | * fixing. |
| 192 | */ |
| 193 | if (b == fix) |
| 194 | { |
| 195 | if (ocfs2_test_bit(i, data)) |
| 196 | ocfs2_clear_bit(i, data); |
| 197 | else |
| 198 | ocfs2_set_bit(i, data); |
| 199 | break; |
| 200 | } |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | void ocfs2_hamming_fix_block(void *data, unsigned int blocksize, |
| 205 | unsigned int fix) |
| 206 | { |
| 207 | ocfs2_hamming_fix(data, blocksize * 8, 0, fix); |
| 208 | } |
| 209 | |
| 210 | /* |
| 211 | * This function generates check information for a block. |
| 212 | * data is the block to be checked. bc is a pointer to the |
| 213 | * ocfs2_block_check structure describing the crc32 and the ecc. |
| 214 | * |
| 215 | * bc should be a pointer inside data, as the function will |
| 216 | * take care of zeroing it before calculating the check information. If |
| 217 | * bc does not point inside data, the caller must make sure any inline |
| 218 | * ocfs2_block_check structures are zeroed. |
| 219 | * |
| 220 | * The data buffer must be in on-disk endian (little endian for ocfs2). |
| 221 | * bc will be filled with little-endian values and will be ready to go to |
| 222 | * disk. |
| 223 | */ |
| 224 | void ocfs2_block_check_compute(void *data, size_t blocksize, |
| 225 | struct ocfs2_block_check *bc) |
| 226 | { |
| 227 | u32 crc; |
| 228 | u32 ecc; |
| 229 | |
| 230 | memset(bc, 0, sizeof(struct ocfs2_block_check)); |
| 231 | |
| 232 | crc = crc32_le(~0, data, blocksize); |
| 233 | ecc = ocfs2_hamming_encode_block(data, blocksize); |
| 234 | |
| 235 | /* |
| 236 | * No ecc'd ocfs2 structure is larger than 4K, so ecc will be no |
| 237 | * larger than 16 bits. |
| 238 | */ |
| 239 | BUG_ON(ecc > USHORT_MAX); |
| 240 | |
| 241 | bc->bc_crc32e = cpu_to_le32(crc); |
| 242 | bc->bc_ecc = cpu_to_le16((u16)ecc); |
| 243 | } |
| 244 | |
| 245 | /* |
| 246 | * This function validates existing check information. Like _compute, |
| 247 | * the function will take care of zeroing bc before calculating check codes. |
| 248 | * If bc is not a pointer inside data, the caller must have zeroed any |
| 249 | * inline ocfs2_block_check structures. |
| 250 | * |
| 251 | * Again, the data passed in should be the on-disk endian. |
| 252 | */ |
| 253 | int ocfs2_block_check_validate(void *data, size_t blocksize, |
| 254 | struct ocfs2_block_check *bc) |
| 255 | { |
| 256 | int rc = 0; |
| 257 | struct ocfs2_block_check check; |
| 258 | u32 crc, ecc; |
| 259 | |
| 260 | check.bc_crc32e = le32_to_cpu(bc->bc_crc32e); |
| 261 | check.bc_ecc = le16_to_cpu(bc->bc_ecc); |
| 262 | |
| 263 | memset(bc, 0, sizeof(struct ocfs2_block_check)); |
| 264 | |
| 265 | /* Fast path - if the crc32 validates, we're good to go */ |
| 266 | crc = crc32_le(~0, data, blocksize); |
| 267 | if (crc == check.bc_crc32e) |
| 268 | goto out; |
| 269 | |
Joel Becker | d6b32bb | 2008-10-17 14:55:01 -0700 | [diff] [blame] | 270 | mlog(ML_ERROR, |
| 271 | "CRC32 failed: stored: %u, computed %u. Applying ECC.\n", |
| 272 | (unsigned int)check.bc_crc32e, (unsigned int)crc); |
| 273 | |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 274 | /* Ok, try ECC fixups */ |
| 275 | ecc = ocfs2_hamming_encode_block(data, blocksize); |
| 276 | ocfs2_hamming_fix_block(data, blocksize, ecc ^ check.bc_ecc); |
| 277 | |
| 278 | /* And check the crc32 again */ |
| 279 | crc = crc32_le(~0, data, blocksize); |
| 280 | if (crc == check.bc_crc32e) |
| 281 | goto out; |
| 282 | |
Joel Becker | d6b32bb | 2008-10-17 14:55:01 -0700 | [diff] [blame] | 283 | mlog(ML_ERROR, "Fixed CRC32 failed: stored: %u, computed %u\n", |
| 284 | (unsigned int)check.bc_crc32e, (unsigned int)crc); |
| 285 | |
Joel Becker | 70ad1ba | 2008-10-16 17:54:25 -0700 | [diff] [blame] | 286 | rc = -EIO; |
| 287 | |
| 288 | out: |
| 289 | bc->bc_crc32e = cpu_to_le32(check.bc_crc32e); |
| 290 | bc->bc_ecc = cpu_to_le16(check.bc_ecc); |
| 291 | |
| 292 | return rc; |
| 293 | } |
| 294 | |
| 295 | /* |
| 296 | * This function generates check information for a list of buffer_heads. |
| 297 | * bhs is the blocks to be checked. bc is a pointer to the |
| 298 | * ocfs2_block_check structure describing the crc32 and the ecc. |
| 299 | * |
| 300 | * bc should be a pointer inside data, as the function will |
| 301 | * take care of zeroing it before calculating the check information. If |
| 302 | * bc does not point inside data, the caller must make sure any inline |
| 303 | * ocfs2_block_check structures are zeroed. |
| 304 | * |
| 305 | * The data buffer must be in on-disk endian (little endian for ocfs2). |
| 306 | * bc will be filled with little-endian values and will be ready to go to |
| 307 | * disk. |
| 308 | */ |
| 309 | void ocfs2_block_check_compute_bhs(struct buffer_head **bhs, int nr, |
| 310 | struct ocfs2_block_check *bc) |
| 311 | { |
| 312 | int i; |
| 313 | u32 crc, ecc; |
| 314 | |
| 315 | BUG_ON(nr < 0); |
| 316 | |
| 317 | if (!nr) |
| 318 | return; |
| 319 | |
| 320 | memset(bc, 0, sizeof(struct ocfs2_block_check)); |
| 321 | |
| 322 | for (i = 0, crc = ~0, ecc = 0; i < nr; i++) { |
| 323 | crc = crc32_le(crc, bhs[i]->b_data, bhs[i]->b_size); |
| 324 | /* |
| 325 | * The number of bits in a buffer is obviously b_size*8. |
| 326 | * The offset of this buffer is b_size*i, so the bit offset |
| 327 | * of this buffer is b_size*8*i. |
| 328 | */ |
| 329 | ecc = (u16)ocfs2_hamming_encode(ecc, bhs[i]->b_data, |
| 330 | bhs[i]->b_size * 8, |
| 331 | bhs[i]->b_size * 8 * i); |
| 332 | } |
| 333 | |
| 334 | /* |
| 335 | * No ecc'd ocfs2 structure is larger than 4K, so ecc will be no |
| 336 | * larger than 16 bits. |
| 337 | */ |
| 338 | BUG_ON(ecc > USHORT_MAX); |
| 339 | |
| 340 | bc->bc_crc32e = cpu_to_le32(crc); |
| 341 | bc->bc_ecc = cpu_to_le16((u16)ecc); |
| 342 | } |
| 343 | |
| 344 | /* |
| 345 | * This function validates existing check information on a list of |
| 346 | * buffer_heads. Like _compute_bhs, the function will take care of |
| 347 | * zeroing bc before calculating check codes. If bc is not a pointer |
| 348 | * inside data, the caller must have zeroed any inline |
| 349 | * ocfs2_block_check structures. |
| 350 | * |
| 351 | * Again, the data passed in should be the on-disk endian. |
| 352 | */ |
| 353 | int ocfs2_block_check_validate_bhs(struct buffer_head **bhs, int nr, |
| 354 | struct ocfs2_block_check *bc) |
| 355 | { |
| 356 | int i, rc = 0; |
| 357 | struct ocfs2_block_check check; |
| 358 | u32 crc, ecc, fix; |
| 359 | |
| 360 | BUG_ON(nr < 0); |
| 361 | |
| 362 | if (!nr) |
| 363 | return 0; |
| 364 | |
| 365 | check.bc_crc32e = le32_to_cpu(bc->bc_crc32e); |
| 366 | check.bc_ecc = le16_to_cpu(bc->bc_ecc); |
| 367 | |
| 368 | memset(bc, 0, sizeof(struct ocfs2_block_check)); |
| 369 | |
| 370 | /* Fast path - if the crc32 validates, we're good to go */ |
| 371 | for (i = 0, crc = ~0; i < nr; i++) |
| 372 | crc = crc32_le(crc, bhs[i]->b_data, bhs[i]->b_size); |
| 373 | if (crc == check.bc_crc32e) |
| 374 | goto out; |
| 375 | |
| 376 | mlog(ML_ERROR, |
| 377 | "CRC32 failed: stored: %u, computed %u. Applying ECC.\n", |
| 378 | (unsigned int)check.bc_crc32e, (unsigned int)crc); |
| 379 | |
| 380 | /* Ok, try ECC fixups */ |
| 381 | for (i = 0, ecc = 0; i < nr; i++) { |
| 382 | /* |
| 383 | * The number of bits in a buffer is obviously b_size*8. |
| 384 | * The offset of this buffer is b_size*i, so the bit offset |
| 385 | * of this buffer is b_size*8*i. |
| 386 | */ |
| 387 | ecc = (u16)ocfs2_hamming_encode(ecc, bhs[i]->b_data, |
| 388 | bhs[i]->b_size * 8, |
| 389 | bhs[i]->b_size * 8 * i); |
| 390 | } |
| 391 | fix = ecc ^ check.bc_ecc; |
| 392 | for (i = 0; i < nr; i++) { |
| 393 | /* |
| 394 | * Try the fix against each buffer. It will only affect |
| 395 | * one of them. |
| 396 | */ |
| 397 | ocfs2_hamming_fix(bhs[i]->b_data, bhs[i]->b_size * 8, |
| 398 | bhs[i]->b_size * 8 * i, fix); |
| 399 | } |
| 400 | |
| 401 | /* And check the crc32 again */ |
| 402 | for (i = 0, crc = ~0; i < nr; i++) |
| 403 | crc = crc32_le(crc, bhs[i]->b_data, bhs[i]->b_size); |
| 404 | if (crc == check.bc_crc32e) |
| 405 | goto out; |
| 406 | |
| 407 | mlog(ML_ERROR, "Fixed CRC32 failed: stored: %u, computed %u\n", |
| 408 | (unsigned int)check.bc_crc32e, (unsigned int)crc); |
| 409 | |
| 410 | rc = -EIO; |
| 411 | |
| 412 | out: |
| 413 | bc->bc_crc32e = cpu_to_le32(check.bc_crc32e); |
| 414 | bc->bc_ecc = cpu_to_le16(check.bc_ecc); |
| 415 | |
| 416 | return rc; |
| 417 | } |
| 418 | |
| 419 | /* |
| 420 | * These are the main API. They check the superblock flag before |
| 421 | * calling the underlying operations. |
| 422 | * |
| 423 | * They expect the buffer(s) to be in disk format. |
| 424 | */ |
| 425 | void ocfs2_compute_meta_ecc(struct super_block *sb, void *data, |
| 426 | struct ocfs2_block_check *bc) |
| 427 | { |
| 428 | if (ocfs2_meta_ecc(OCFS2_SB(sb))) |
| 429 | ocfs2_block_check_compute(data, sb->s_blocksize, bc); |
| 430 | } |
| 431 | |
| 432 | int ocfs2_validate_meta_ecc(struct super_block *sb, void *data, |
| 433 | struct ocfs2_block_check *bc) |
| 434 | { |
| 435 | int rc = 0; |
| 436 | |
| 437 | if (ocfs2_meta_ecc(OCFS2_SB(sb))) |
| 438 | rc = ocfs2_block_check_validate(data, sb->s_blocksize, bc); |
| 439 | |
| 440 | return rc; |
| 441 | } |
| 442 | |
| 443 | void ocfs2_compute_meta_ecc_bhs(struct super_block *sb, |
| 444 | struct buffer_head **bhs, int nr, |
| 445 | struct ocfs2_block_check *bc) |
| 446 | { |
| 447 | if (ocfs2_meta_ecc(OCFS2_SB(sb))) |
| 448 | ocfs2_block_check_compute_bhs(bhs, nr, bc); |
| 449 | } |
| 450 | |
| 451 | int ocfs2_validate_meta_ecc_bhs(struct super_block *sb, |
| 452 | struct buffer_head **bhs, int nr, |
| 453 | struct ocfs2_block_check *bc) |
| 454 | { |
| 455 | int rc = 0; |
| 456 | |
| 457 | if (ocfs2_meta_ecc(OCFS2_SB(sb))) |
| 458 | rc = ocfs2_block_check_validate_bhs(bhs, nr, bc); |
| 459 | |
| 460 | return rc; |
| 461 | } |
| 462 | |