1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * Checksum and ECC codes for the OCFS2 userspace library.
8 * Copyright (C) 2006, 2008 Oracle. All rights reserved.
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.
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.
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 <linux/debugfs.h>
26 #include <linux/module.h>
28 #include <asm/byteorder.h>
30 #include <cluster/masklog.h>
34 #include "blockcheck.h"
38 * We use the following conventions:
42 * c = # total code bits (d + p)
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
50 * number are offset by all the parity bits beforehand.
52 * Recall that bit numbers in hamming code are 1-based. This function
53 * takes the 0-based data bit from the caller.
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
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.
64 static unsigned int calc_code_bit(unsigned int i
, unsigned int *p_cache
)
66 unsigned int b
, p
= 0;
69 * Data bits are 0-based, but we're talking code bits, which
74 /* Use the cache if it is there */
80 * For every power of two below our bit number, bump our bit.
82 * We compare with (b + 1) because we have to compare with what b
83 * would be _if_ it were bumped up by the parity bit. Capice?
87 for (; (1 << p
) < (b
+ 1); p
++)
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:
102 * parity = ocfs2_hamming_encode(0, buf1, 512 * 8, 0);
103 * parity = ocfs2_hamming_encode(parity, buf2, 512 * 8, 512 * 8);
105 * If you just have one buffer, use ocfs2_hamming_encode_block().
107 u32
ocfs2_hamming_encode(u32 parity
, void *data
, unsigned int d
, unsigned int nr
)
109 unsigned int i
, b
, p
= 0;
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
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
122 for (i
= 0; (i
= ocfs2_find_next_bit(data
, d
, i
)) < d
; i
++)
125 * i is the offset in this hunk, nr + i is the total bit
128 b
= calc_code_bit(nr
+ i
, &p
);
131 * Data bits in the resultant code are checked by
132 * parity bits that are part of the bit number
133 * representation. Huh?
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.
143 * Note that 'k' is the _code_ bit number. 'b' in
149 /* While the data buffer was treated as little endian, the
150 * return value is in host endian. */
154 u32
ocfs2_hamming_encode_block(void *data
, unsigned int blocksize
)
156 return ocfs2_hamming_encode(0, data
, blocksize
* 8, 0);
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.
164 * If you only have one hunk, use ocfs2_hamming_fix_block().
166 void ocfs2_hamming_fix(void *data
, unsigned int d
, unsigned int nr
,
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.
177 if (hweight32(fix
) == 1)
181 * nr + d is the bit right past the data hunk we're looking at.
182 * If fix after that, nothing to do
184 if (fix
>= calc_code_bit(nr
+ d
, NULL
))
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'.
192 b
= calc_code_bit(nr
, NULL
);
193 /* If the fix is before this hunk, nothing to do */
197 for (i
= 0; i
< d
; i
++, b
++)
199 /* Skip past parity bits */
200 while (hweight32(b
) == 1)
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.
208 * Thus, when b == fix, bit i in the current hunk needs
213 if (ocfs2_test_bit(i
, data
))
214 ocfs2_clear_bit(i
, data
);
216 ocfs2_set_bit(i
, data
);
222 void ocfs2_hamming_fix_block(void *data
, unsigned int blocksize
,
225 ocfs2_hamming_fix(data
, blocksize
* 8, 0, fix
);
233 #ifdef CONFIG_DEBUG_FS
235 static int blockcheck_u64_get(void *data
, u64
*val
)
240 DEFINE_SIMPLE_ATTRIBUTE(blockcheck_fops
, blockcheck_u64_get
, NULL
, "%llu\n");
242 static struct dentry
*blockcheck_debugfs_create(const char *name
,
243 struct dentry
*parent
,
246 return debugfs_create_file(name
, S_IFREG
| S_IRUSR
, parent
, value
,
250 static void ocfs2_blockcheck_debug_remove(struct ocfs2_blockcheck_stats
*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
;
264 static int ocfs2_blockcheck_debug_install(struct ocfs2_blockcheck_stats
*stats
,
265 struct dentry
*parent
)
272 stats
->b_debug_dir
= debugfs_create_dir("blockcheck", parent
);
273 if (!stats
->b_debug_dir
)
276 stats
->b_debug_check
=
277 blockcheck_debugfs_create("blocks_checked",
279 &stats
->b_check_count
);
281 stats
->b_debug_failure
=
282 blockcheck_debugfs_create("checksums_failed",
284 &stats
->b_failure_count
);
286 stats
->b_debug_recover
=
287 blockcheck_debugfs_create("ecc_recoveries",
289 &stats
->b_recover_count
);
290 if (stats
->b_debug_check
&& stats
->b_debug_failure
&&
291 stats
->b_debug_recover
)
296 ocfs2_blockcheck_debug_remove(stats
);
300 static inline int ocfs2_blockcheck_debug_install(struct ocfs2_blockcheck_stats
*stats
,
301 struct dentry
*parent
)
306 static inline void ocfs2_blockcheck_debug_remove(struct ocfs2_blockcheck_stats
*stats
)
309 #endif /* CONFIG_DEBUG_FS */
311 /* Always-called wrappers for starting and stopping the debugfs files */
312 int ocfs2_blockcheck_stats_debugfs_install(struct ocfs2_blockcheck_stats
*stats
,
313 struct dentry
*parent
)
315 return ocfs2_blockcheck_debug_install(stats
, parent
);
318 void ocfs2_blockcheck_stats_debugfs_remove(struct ocfs2_blockcheck_stats
*stats
)
320 ocfs2_blockcheck_debug_remove(stats
);
323 static void ocfs2_blockcheck_inc_check(struct ocfs2_blockcheck_stats
*stats
)
330 spin_lock(&stats
->b_lock
);
331 stats
->b_check_count
++;
332 new_count
= stats
->b_check_count
;
333 spin_unlock(&stats
->b_lock
);
336 mlog(ML_NOTICE
, "Block check count has wrapped\n");
339 static void ocfs2_blockcheck_inc_failure(struct ocfs2_blockcheck_stats
*stats
)
346 spin_lock(&stats
->b_lock
);
347 stats
->b_failure_count
++;
348 new_count
= stats
->b_failure_count
;
349 spin_unlock(&stats
->b_lock
);
352 mlog(ML_NOTICE
, "Checksum failure count has wrapped\n");
355 static void ocfs2_blockcheck_inc_recover(struct ocfs2_blockcheck_stats
*stats
)
362 spin_lock(&stats
->b_lock
);
363 stats
->b_recover_count
++;
364 new_count
= stats
->b_recover_count
;
365 spin_unlock(&stats
->b_lock
);
368 mlog(ML_NOTICE
, "ECC recovery count has wrapped\n");
374 * These are the low-level APIs for using the ocfs2_block_check structure.
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.
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.
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
391 void ocfs2_block_check_compute(void *data
, size_t blocksize
,
392 struct ocfs2_block_check
*bc
)
397 memset(bc
, 0, sizeof(struct ocfs2_block_check
));
399 crc
= crc32_le(~0, data
, blocksize
);
400 ecc
= ocfs2_hamming_encode_block(data
, blocksize
);
403 * No ecc'd ocfs2 structure is larger than 4K, so ecc will be no
404 * larger than 16 bits.
406 BUG_ON(ecc
> USHORT_MAX
);
408 bc
->bc_crc32e
= cpu_to_le32(crc
);
409 bc
->bc_ecc
= cpu_to_le16((u16
)ecc
);
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.
418 * Again, the data passed in should be the on-disk endian.
420 int ocfs2_block_check_validate(void *data
, size_t blocksize
,
421 struct ocfs2_block_check
*bc
,
422 struct ocfs2_blockcheck_stats
*stats
)
425 struct ocfs2_block_check check
;
428 ocfs2_blockcheck_inc_check(stats
);
430 check
.bc_crc32e
= le32_to_cpu(bc
->bc_crc32e
);
431 check
.bc_ecc
= le16_to_cpu(bc
->bc_ecc
);
433 memset(bc
, 0, sizeof(struct ocfs2_block_check
));
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
)
440 ocfs2_blockcheck_inc_failure(stats
);
442 "CRC32 failed: stored: %u, computed %u. Applying ECC.\n",
443 (unsigned int)check
.bc_crc32e
, (unsigned int)crc
);
445 /* Ok, try ECC fixups */
446 ecc
= ocfs2_hamming_encode_block(data
, blocksize
);
447 ocfs2_hamming_fix_block(data
, blocksize
, ecc
^ check
.bc_ecc
);
449 /* And check the crc32 again */
450 crc
= crc32_le(~0, data
, blocksize
);
451 if (crc
== check
.bc_crc32e
) {
452 ocfs2_blockcheck_inc_recover(stats
);
456 mlog(ML_ERROR
, "Fixed CRC32 failed: stored: %u, computed %u\n",
457 (unsigned int)check
.bc_crc32e
, (unsigned int)crc
);
462 bc
->bc_crc32e
= cpu_to_le32(check
.bc_crc32e
);
463 bc
->bc_ecc
= cpu_to_le16(check
.bc_ecc
);
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.
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.
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
482 void ocfs2_block_check_compute_bhs(struct buffer_head
**bhs
, int nr
,
483 struct ocfs2_block_check
*bc
)
493 memset(bc
, 0, sizeof(struct ocfs2_block_check
));
495 for (i
= 0, crc
= ~0, ecc
= 0; i
< nr
; i
++) {
496 crc
= crc32_le(crc
, bhs
[i
]->b_data
, bhs
[i
]->b_size
);
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.
502 ecc
= (u16
)ocfs2_hamming_encode(ecc
, bhs
[i
]->b_data
,
504 bhs
[i
]->b_size
* 8 * i
);
508 * No ecc'd ocfs2 structure is larger than 4K, so ecc will be no
509 * larger than 16 bits.
511 BUG_ON(ecc
> USHORT_MAX
);
513 bc
->bc_crc32e
= cpu_to_le32(crc
);
514 bc
->bc_ecc
= cpu_to_le16((u16
)ecc
);
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.
524 * Again, the data passed in should be the on-disk endian.
526 int ocfs2_block_check_validate_bhs(struct buffer_head
**bhs
, int nr
,
527 struct ocfs2_block_check
*bc
,
528 struct ocfs2_blockcheck_stats
*stats
)
531 struct ocfs2_block_check check
;
539 ocfs2_blockcheck_inc_check(stats
);
541 check
.bc_crc32e
= le32_to_cpu(bc
->bc_crc32e
);
542 check
.bc_ecc
= le16_to_cpu(bc
->bc_ecc
);
544 memset(bc
, 0, sizeof(struct ocfs2_block_check
));
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
)
552 ocfs2_blockcheck_inc_failure(stats
);
554 "CRC32 failed: stored: %u, computed %u. Applying ECC.\n",
555 (unsigned int)check
.bc_crc32e
, (unsigned int)crc
);
557 /* Ok, try ECC fixups */
558 for (i
= 0, ecc
= 0; i
< nr
; i
++) {
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.
564 ecc
= (u16
)ocfs2_hamming_encode(ecc
, bhs
[i
]->b_data
,
566 bhs
[i
]->b_size
* 8 * i
);
568 fix
= ecc
^ check
.bc_ecc
;
569 for (i
= 0; i
< nr
; i
++) {
571 * Try the fix against each buffer. It will only affect
574 ocfs2_hamming_fix(bhs
[i
]->b_data
, bhs
[i
]->b_size
* 8,
575 bhs
[i
]->b_size
* 8 * i
, fix
);
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
);
581 if (crc
== check
.bc_crc32e
) {
582 ocfs2_blockcheck_inc_recover(stats
);
586 mlog(ML_ERROR
, "Fixed CRC32 failed: stored: %u, computed %u\n",
587 (unsigned int)check
.bc_crc32e
, (unsigned int)crc
);
592 bc
->bc_crc32e
= cpu_to_le32(check
.bc_crc32e
);
593 bc
->bc_ecc
= cpu_to_le16(check
.bc_ecc
);
599 * These are the main API. They check the superblock flag before
600 * calling the underlying operations.
602 * They expect the buffer(s) to be in disk format.
604 void ocfs2_compute_meta_ecc(struct super_block
*sb
, void *data
,
605 struct ocfs2_block_check
*bc
)
607 if (ocfs2_meta_ecc(OCFS2_SB(sb
)))
608 ocfs2_block_check_compute(data
, sb
->s_blocksize
, bc
);
611 int ocfs2_validate_meta_ecc(struct super_block
*sb
, void *data
,
612 struct ocfs2_block_check
*bc
)
615 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
617 if (ocfs2_meta_ecc(osb
))
618 rc
= ocfs2_block_check_validate(data
, sb
->s_blocksize
, bc
,
619 &osb
->osb_ecc_stats
);
624 void ocfs2_compute_meta_ecc_bhs(struct super_block
*sb
,
625 struct buffer_head
**bhs
, int nr
,
626 struct ocfs2_block_check
*bc
)
628 if (ocfs2_meta_ecc(OCFS2_SB(sb
)))
629 ocfs2_block_check_compute_bhs(bhs
, nr
, bc
);
632 int ocfs2_validate_meta_ecc_bhs(struct super_block
*sb
,
633 struct buffer_head
**bhs
, int nr
,
634 struct ocfs2_block_check
*bc
)
637 struct ocfs2_super
*osb
= OCFS2_SB(sb
);
639 if (ocfs2_meta_ecc(osb
))
640 rc
= ocfs2_block_check_validate_bhs(bhs
, nr
, bc
,
641 &osb
->osb_ecc_stats
);