]> bbs.cooldavid.org Git - net-next-2.6.git/blame - fs/ocfs2/blockcheck.c
xps: Transmit Packet Steering
[net-next-2.6.git] / fs / ocfs2 / blockcheck.c
CommitLineData
70ad1ba7
JB
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>
73be192b
JB
25#include <linux/debugfs.h>
26#include <linux/module.h>
27#include <linux/fs.h>
70ad1ba7
JB
28#include <asm/byteorder.h>
29
d6b32bbb
JB
30#include <cluster/masklog.h>
31
70ad1ba7
JB
32#include "ocfs2.h"
33
34#include "blockcheck.h"
35
36
70ad1ba7
JB
37/*
38 * We use the following conventions:
39 *
40 * d = # data bits
41 * p = # parity bits
42 * c = # total code bits (d + p)
43 */
70ad1ba7 44
7bb458a5 45
70ad1ba7
JB
46/*
47 * Calculate the bit offset in the hamming code buffer based on the bit's
48 * offset in the data buffer. Since the hamming code reserves all
49 * power-of-two bits for parity, the data bit number and the code bit
bf48aabb 50 * number are offset by all the parity bits beforehand.
70ad1ba7
JB
51 *
52 * Recall that bit numbers in hamming code are 1-based. This function
53 * takes the 0-based data bit from the caller.
54 *
55 * An example. Take bit 1 of the data buffer. 1 is a power of two (2^0),
56 * so it's a parity bit. 2 is a power of two (2^1), so it's a parity bit.
57 * 3 is not a power of two. So bit 1 of the data buffer ends up as bit 3
58 * in the code buffer.
58896c4d
JB
59 *
60 * The caller can pass in *p if it wants to keep track of the most recent
61 * number of parity bits added. This allows the function to start the
62 * calculation at the last place.
70ad1ba7 63 */
58896c4d 64static unsigned int calc_code_bit(unsigned int i, unsigned int *p_cache)
70ad1ba7 65{
58896c4d 66 unsigned int b, p = 0;
70ad1ba7
JB
67
68 /*
69 * Data bits are 0-based, but we're talking code bits, which
70 * are 1-based.
71 */
72 b = i + 1;
73
58896c4d
JB
74 /* Use the cache if it is there */
75 if (p_cache)
76 p = *p_cache;
7bb458a5
JB
77 b += p;
78
70ad1ba7
JB
79 /*
80 * For every power of two below our bit number, bump our bit.
81 *
58896c4d 82 * We compare with (b + 1) because we have to compare with what b
70ad1ba7 83 * would be _if_ it were bumped up by the parity bit. Capice?
7bb458a5 84 *
58896c4d 85 * p is set above.
70ad1ba7 86 */
58896c4d 87 for (; (1 << p) < (b + 1); p++)
70ad1ba7
JB
88 b++;
89
58896c4d
JB
90 if (p_cache)
91 *p_cache = p;
92
70ad1ba7
JB
93 return b;
94}
95
96/*
97 * This is the low level encoder function. It can be called across
98 * multiple hunks just like the crc32 code. 'd' is the number of bits
99 * _in_this_hunk_. nr is the bit offset of this hunk. So, if you had
100 * two 512B buffers, you would do it like so:
101 *
102 * parity = ocfs2_hamming_encode(0, buf1, 512 * 8, 0);
103 * parity = ocfs2_hamming_encode(parity, buf2, 512 * 8, 512 * 8);
104 *
105 * If you just have one buffer, use ocfs2_hamming_encode_block().
106 */
107u32 ocfs2_hamming_encode(u32 parity, void *data, unsigned int d, unsigned int nr)
108{
58896c4d 109 unsigned int i, b, p = 0;
70ad1ba7 110
e798b3f8 111 BUG_ON(!d);
70ad1ba7
JB
112
113 /*
114 * b is the hamming code bit number. Hamming code specifies a
115 * 1-based array, but C uses 0-based. So 'i' is for C, and 'b' is
116 * for the algorithm.
117 *
118 * The i++ in the for loop is so that the start offset passed
119 * to ocfs2_find_next_bit_set() is one greater than the previously
120 * found bit.
121 */
122 for (i = 0; (i = ocfs2_find_next_bit(data, d, i)) < d; i++)
123 {
124 /*
125 * i is the offset in this hunk, nr + i is the total bit
126 * offset.
127 */
58896c4d 128 b = calc_code_bit(nr + i, &p);
70ad1ba7 129
e798b3f8
JB
130 /*
131 * Data bits in the resultant code are checked by
132 * parity bits that are part of the bit number
133 * representation. Huh?
134 *
135 * <wikipedia href="http://en.wikipedia.org/wiki/Hamming_code">
136 * In other words, the parity bit at position 2^k
137 * checks bits in positions having bit k set in
138 * their binary representation. Conversely, for
139 * instance, bit 13, i.e. 1101(2), is checked by
140 * bits 1000(2) = 8, 0100(2)=4 and 0001(2) = 1.
141 * </wikipedia>
142 *
143 * Note that 'k' is the _code_ bit number. 'b' in
144 * our loop.
145 */
146 parity ^= b;
70ad1ba7
JB
147 }
148
149 /* While the data buffer was treated as little endian, the
150 * return value is in host endian. */
151 return parity;
152}
153
154u32 ocfs2_hamming_encode_block(void *data, unsigned int blocksize)
155{
156 return ocfs2_hamming_encode(0, data, blocksize * 8, 0);
157}
158
159/*
160 * Like ocfs2_hamming_encode(), this can handle hunks. nr is the bit
161 * offset of the current hunk. If bit to be fixed is not part of the
162 * current hunk, this does nothing.
163 *
164 * If you only have one hunk, use ocfs2_hamming_fix_block().
165 */
166void ocfs2_hamming_fix(void *data, unsigned int d, unsigned int nr,
167 unsigned int fix)
168{
70ad1ba7
JB
169 unsigned int i, b;
170
e798b3f8 171 BUG_ON(!d);
70ad1ba7
JB
172
173 /*
174 * If the bit to fix has an hweight of 1, it's a parity bit. One
175 * busted parity bit is its own error. Nothing to do here.
176 */
177 if (hweight32(fix) == 1)
178 return;
179
180 /*
181 * nr + d is the bit right past the data hunk we're looking at.
182 * If fix after that, nothing to do
183 */
58896c4d 184 if (fix >= calc_code_bit(nr + d, NULL))
70ad1ba7
JB
185 return;
186
187 /*
188 * nr is the offset in the data hunk we're starting at. Let's
189 * start b at the offset in the code buffer. See hamming_encode()
190 * for a more detailed description of 'b'.
191 */
58896c4d 192 b = calc_code_bit(nr, NULL);
70ad1ba7
JB
193 /* If the fix is before this hunk, nothing to do */
194 if (fix < b)
195 return;
196
197 for (i = 0; i < d; i++, b++)
198 {
199 /* Skip past parity bits */
200 while (hweight32(b) == 1)
201 b++;
202
203 /*
204 * i is the offset in this data hunk.
205 * nr + i is the offset in the total data buffer.
206 * b is the offset in the total code buffer.
207 *
208 * Thus, when b == fix, bit i in the current hunk needs
209 * fixing.
210 */
211 if (b == fix)
212 {
213 if (ocfs2_test_bit(i, data))
214 ocfs2_clear_bit(i, data);
215 else
216 ocfs2_set_bit(i, data);
217 break;
218 }
219 }
220}
221
222void ocfs2_hamming_fix_block(void *data, unsigned int blocksize,
223 unsigned int fix)
224{
225 ocfs2_hamming_fix(data, blocksize * 8, 0, fix);
226}
227
73be192b
JB
228
229/*
230 * Debugfs handling.
231 */
232
233#ifdef CONFIG_DEBUG_FS
234
235static int blockcheck_u64_get(void *data, u64 *val)
236{
237 *val = *(u64 *)data;
238 return 0;
239}
240DEFINE_SIMPLE_ATTRIBUTE(blockcheck_fops, blockcheck_u64_get, NULL, "%llu\n");
241
242static struct dentry *blockcheck_debugfs_create(const char *name,
243 struct dentry *parent,
244 u64 *value)
245{
246 return debugfs_create_file(name, S_IFREG | S_IRUSR, parent, value,
247 &blockcheck_fops);
248}
249
250static void ocfs2_blockcheck_debug_remove(struct ocfs2_blockcheck_stats *stats)
251{
252 if (stats) {
253 debugfs_remove(stats->b_debug_check);
254 stats->b_debug_check = NULL;
255 debugfs_remove(stats->b_debug_failure);
256 stats->b_debug_failure = NULL;
257 debugfs_remove(stats->b_debug_recover);
258 stats->b_debug_recover = NULL;
259 debugfs_remove(stats->b_debug_dir);
260 stats->b_debug_dir = NULL;
261 }
262}
263
264static int ocfs2_blockcheck_debug_install(struct ocfs2_blockcheck_stats *stats,
265 struct dentry *parent)
266{
267 int rc = -EINVAL;
268
269 if (!stats)
270 goto out;
271
272 stats->b_debug_dir = debugfs_create_dir("blockcheck", parent);
273 if (!stats->b_debug_dir)
274 goto out;
275
276 stats->b_debug_check =
277 blockcheck_debugfs_create("blocks_checked",
278 stats->b_debug_dir,
279 &stats->b_check_count);
280
281 stats->b_debug_failure =
282 blockcheck_debugfs_create("checksums_failed",
283 stats->b_debug_dir,
284 &stats->b_failure_count);
285
286 stats->b_debug_recover =
287 blockcheck_debugfs_create("ecc_recoveries",
288 stats->b_debug_dir,
289 &stats->b_recover_count);
290 if (stats->b_debug_check && stats->b_debug_failure &&
291 stats->b_debug_recover)
292 rc = 0;
293
294out:
295 if (rc)
296 ocfs2_blockcheck_debug_remove(stats);
297 return rc;
298}
299#else
300static inline int ocfs2_blockcheck_debug_install(struct ocfs2_blockcheck_stats *stats,
301 struct dentry *parent)
302{
303 return 0;
304}
305
306static inline void ocfs2_blockcheck_debug_remove(struct ocfs2_blockcheck_stats *stats)
307{
308}
309#endif /* CONFIG_DEBUG_FS */
310
311/* Always-called wrappers for starting and stopping the debugfs files */
312int ocfs2_blockcheck_stats_debugfs_install(struct ocfs2_blockcheck_stats *stats,
313 struct dentry *parent)
314{
315 return ocfs2_blockcheck_debug_install(stats, parent);
316}
317
318void ocfs2_blockcheck_stats_debugfs_remove(struct ocfs2_blockcheck_stats *stats)
319{
320 ocfs2_blockcheck_debug_remove(stats);
321}
322
323static void ocfs2_blockcheck_inc_check(struct ocfs2_blockcheck_stats *stats)
324{
325 u64 new_count;
326
327 if (!stats)
328 return;
329
330 spin_lock(&stats->b_lock);
331 stats->b_check_count++;
332 new_count = stats->b_check_count;
333 spin_unlock(&stats->b_lock);
334
335 if (!new_count)
336 mlog(ML_NOTICE, "Block check count has wrapped\n");
337}
338
339static void ocfs2_blockcheck_inc_failure(struct ocfs2_blockcheck_stats *stats)
340{
341 u64 new_count;
342
343 if (!stats)
344 return;
345
346 spin_lock(&stats->b_lock);
347 stats->b_failure_count++;
348 new_count = stats->b_failure_count;
349 spin_unlock(&stats->b_lock);
350
351 if (!new_count)
352 mlog(ML_NOTICE, "Checksum failure count has wrapped\n");
353}
354
355static void ocfs2_blockcheck_inc_recover(struct ocfs2_blockcheck_stats *stats)
356{
357 u64 new_count;
358
359 if (!stats)
360 return;
361
362 spin_lock(&stats->b_lock);
363 stats->b_recover_count++;
364 new_count = stats->b_recover_count;
365 spin_unlock(&stats->b_lock);
366
367 if (!new_count)
368 mlog(ML_NOTICE, "ECC recovery count has wrapped\n");
369}
370
371
372
373/*
374 * These are the low-level APIs for using the ocfs2_block_check structure.
375 */
376
70ad1ba7
JB
377/*
378 * This function generates check information for a block.
379 * data is the block to be checked. bc is a pointer to the
380 * ocfs2_block_check structure describing the crc32 and the ecc.
381 *
382 * bc should be a pointer inside data, as the function will
383 * take care of zeroing it before calculating the check information. If
384 * bc does not point inside data, the caller must make sure any inline
385 * ocfs2_block_check structures are zeroed.
386 *
387 * The data buffer must be in on-disk endian (little endian for ocfs2).
388 * bc will be filled with little-endian values and will be ready to go to
389 * disk.
390 */
391void ocfs2_block_check_compute(void *data, size_t blocksize,
392 struct ocfs2_block_check *bc)
393{
394 u32 crc;
395 u32 ecc;
396
397 memset(bc, 0, sizeof(struct ocfs2_block_check));
398
399 crc = crc32_le(~0, data, blocksize);
400 ecc = ocfs2_hamming_encode_block(data, blocksize);
401
402 /*
403 * No ecc'd ocfs2 structure is larger than 4K, so ecc will be no
404 * larger than 16 bits.
405 */
4be929be 406 BUG_ON(ecc > USHRT_MAX);
70ad1ba7
JB
407
408 bc->bc_crc32e = cpu_to_le32(crc);
409 bc->bc_ecc = cpu_to_le16((u16)ecc);
410}
411
412/*
413 * This function validates existing check information. Like _compute,
414 * the function will take care of zeroing bc before calculating check codes.
415 * If bc is not a pointer inside data, the caller must have zeroed any
416 * inline ocfs2_block_check structures.
417 *
418 * Again, the data passed in should be the on-disk endian.
419 */
420int ocfs2_block_check_validate(void *data, size_t blocksize,
73be192b
JB
421 struct ocfs2_block_check *bc,
422 struct ocfs2_blockcheck_stats *stats)
70ad1ba7
JB
423{
424 int rc = 0;
425 struct ocfs2_block_check check;
426 u32 crc, ecc;
427
73be192b
JB
428 ocfs2_blockcheck_inc_check(stats);
429
70ad1ba7
JB
430 check.bc_crc32e = le32_to_cpu(bc->bc_crc32e);
431 check.bc_ecc = le16_to_cpu(bc->bc_ecc);
432
433 memset(bc, 0, sizeof(struct ocfs2_block_check));
434
435 /* Fast path - if the crc32 validates, we're good to go */
436 crc = crc32_le(~0, data, blocksize);
437 if (crc == check.bc_crc32e)
438 goto out;
439
73be192b 440 ocfs2_blockcheck_inc_failure(stats);
d6b32bbb 441 mlog(ML_ERROR,
dc696ace 442 "CRC32 failed: stored: 0x%x, computed 0x%x. Applying ECC.\n",
d6b32bbb
JB
443 (unsigned int)check.bc_crc32e, (unsigned int)crc);
444
70ad1ba7
JB
445 /* Ok, try ECC fixups */
446 ecc = ocfs2_hamming_encode_block(data, blocksize);
447 ocfs2_hamming_fix_block(data, blocksize, ecc ^ check.bc_ecc);
448
449 /* And check the crc32 again */
450 crc = crc32_le(~0, data, blocksize);
73be192b
JB
451 if (crc == check.bc_crc32e) {
452 ocfs2_blockcheck_inc_recover(stats);
70ad1ba7 453 goto out;
73be192b 454 }
70ad1ba7 455
dc696ace 456 mlog(ML_ERROR, "Fixed CRC32 failed: stored: 0x%x, computed 0x%x\n",
d6b32bbb
JB
457 (unsigned int)check.bc_crc32e, (unsigned int)crc);
458
70ad1ba7
JB
459 rc = -EIO;
460
461out:
462 bc->bc_crc32e = cpu_to_le32(check.bc_crc32e);
463 bc->bc_ecc = cpu_to_le16(check.bc_ecc);
464
465 return rc;
466}
467
468/*
469 * This function generates check information for a list of buffer_heads.
470 * bhs is the blocks to be checked. bc is a pointer to the
471 * ocfs2_block_check structure describing the crc32 and the ecc.
472 *
473 * bc should be a pointer inside data, as the function will
474 * take care of zeroing it before calculating the check information. If
475 * bc does not point inside data, the caller must make sure any inline
476 * ocfs2_block_check structures are zeroed.
477 *
478 * The data buffer must be in on-disk endian (little endian for ocfs2).
479 * bc will be filled with little-endian values and will be ready to go to
480 * disk.
481 */
482void ocfs2_block_check_compute_bhs(struct buffer_head **bhs, int nr,
483 struct ocfs2_block_check *bc)
484{
485 int i;
486 u32 crc, ecc;
487
488 BUG_ON(nr < 0);
489
490 if (!nr)
491 return;
492
493 memset(bc, 0, sizeof(struct ocfs2_block_check));
494
495 for (i = 0, crc = ~0, ecc = 0; i < nr; i++) {
496 crc = crc32_le(crc, bhs[i]->b_data, bhs[i]->b_size);
497 /*
498 * The number of bits in a buffer is obviously b_size*8.
499 * The offset of this buffer is b_size*i, so the bit offset
500 * of this buffer is b_size*8*i.
501 */
502 ecc = (u16)ocfs2_hamming_encode(ecc, bhs[i]->b_data,
503 bhs[i]->b_size * 8,
504 bhs[i]->b_size * 8 * i);
505 }
506
507 /*
508 * No ecc'd ocfs2 structure is larger than 4K, so ecc will be no
509 * larger than 16 bits.
510 */
4be929be 511 BUG_ON(ecc > USHRT_MAX);
70ad1ba7
JB
512
513 bc->bc_crc32e = cpu_to_le32(crc);
514 bc->bc_ecc = cpu_to_le16((u16)ecc);
515}
516
517/*
518 * This function validates existing check information on a list of
519 * buffer_heads. Like _compute_bhs, the function will take care of
520 * zeroing bc before calculating check codes. If bc is not a pointer
521 * inside data, the caller must have zeroed any inline
522 * ocfs2_block_check structures.
523 *
524 * Again, the data passed in should be the on-disk endian.
525 */
526int ocfs2_block_check_validate_bhs(struct buffer_head **bhs, int nr,
73be192b
JB
527 struct ocfs2_block_check *bc,
528 struct ocfs2_blockcheck_stats *stats)
70ad1ba7
JB
529{
530 int i, rc = 0;
531 struct ocfs2_block_check check;
532 u32 crc, ecc, fix;
533
534 BUG_ON(nr < 0);
535
536 if (!nr)
537 return 0;
538
73be192b
JB
539 ocfs2_blockcheck_inc_check(stats);
540
70ad1ba7
JB
541 check.bc_crc32e = le32_to_cpu(bc->bc_crc32e);
542 check.bc_ecc = le16_to_cpu(bc->bc_ecc);
543
544 memset(bc, 0, sizeof(struct ocfs2_block_check));
545
546 /* Fast path - if the crc32 validates, we're good to go */
547 for (i = 0, crc = ~0; i < nr; i++)
548 crc = crc32_le(crc, bhs[i]->b_data, bhs[i]->b_size);
549 if (crc == check.bc_crc32e)
550 goto out;
551
73be192b 552 ocfs2_blockcheck_inc_failure(stats);
70ad1ba7
JB
553 mlog(ML_ERROR,
554 "CRC32 failed: stored: %u, computed %u. Applying ECC.\n",
555 (unsigned int)check.bc_crc32e, (unsigned int)crc);
556
557 /* Ok, try ECC fixups */
558 for (i = 0, ecc = 0; i < nr; i++) {
559 /*
560 * The number of bits in a buffer is obviously b_size*8.
561 * The offset of this buffer is b_size*i, so the bit offset
562 * of this buffer is b_size*8*i.
563 */
564 ecc = (u16)ocfs2_hamming_encode(ecc, bhs[i]->b_data,
565 bhs[i]->b_size * 8,
566 bhs[i]->b_size * 8 * i);
567 }
568 fix = ecc ^ check.bc_ecc;
569 for (i = 0; i < nr; i++) {
570 /*
571 * Try the fix against each buffer. It will only affect
572 * one of them.
573 */
574 ocfs2_hamming_fix(bhs[i]->b_data, bhs[i]->b_size * 8,
575 bhs[i]->b_size * 8 * i, fix);
576 }
577
578 /* And check the crc32 again */
579 for (i = 0, crc = ~0; i < nr; i++)
580 crc = crc32_le(crc, bhs[i]->b_data, bhs[i]->b_size);
73be192b
JB
581 if (crc == check.bc_crc32e) {
582 ocfs2_blockcheck_inc_recover(stats);
70ad1ba7 583 goto out;
73be192b 584 }
70ad1ba7
JB
585
586 mlog(ML_ERROR, "Fixed CRC32 failed: stored: %u, computed %u\n",
587 (unsigned int)check.bc_crc32e, (unsigned int)crc);
588
589 rc = -EIO;
590
591out:
592 bc->bc_crc32e = cpu_to_le32(check.bc_crc32e);
593 bc->bc_ecc = cpu_to_le16(check.bc_ecc);
594
595 return rc;
596}
597
598/*
599 * These are the main API. They check the superblock flag before
600 * calling the underlying operations.
601 *
602 * They expect the buffer(s) to be in disk format.
603 */
604void ocfs2_compute_meta_ecc(struct super_block *sb, void *data,
605 struct ocfs2_block_check *bc)
606{
607 if (ocfs2_meta_ecc(OCFS2_SB(sb)))
608 ocfs2_block_check_compute(data, sb->s_blocksize, bc);
609}
610
611int ocfs2_validate_meta_ecc(struct super_block *sb, void *data,
612 struct ocfs2_block_check *bc)
613{
614 int rc = 0;
73be192b 615 struct ocfs2_super *osb = OCFS2_SB(sb);
70ad1ba7 616
73be192b
JB
617 if (ocfs2_meta_ecc(osb))
618 rc = ocfs2_block_check_validate(data, sb->s_blocksize, bc,
619 &osb->osb_ecc_stats);
70ad1ba7
JB
620
621 return rc;
622}
623
624void ocfs2_compute_meta_ecc_bhs(struct super_block *sb,
625 struct buffer_head **bhs, int nr,
626 struct ocfs2_block_check *bc)
627{
628 if (ocfs2_meta_ecc(OCFS2_SB(sb)))
629 ocfs2_block_check_compute_bhs(bhs, nr, bc);
630}
631
632int ocfs2_validate_meta_ecc_bhs(struct super_block *sb,
633 struct buffer_head **bhs, int nr,
634 struct ocfs2_block_check *bc)
635{
636 int rc = 0;
73be192b 637 struct ocfs2_super *osb = OCFS2_SB(sb);
70ad1ba7 638
73be192b
JB
639 if (ocfs2_meta_ecc(osb))
640 rc = ocfs2_block_check_validate_bhs(bhs, nr, bc,
641 &osb->osb_ecc_stats);
70ad1ba7
JB
642
643 return rc;
644}
645