1 Index: linux-2.6.16.53-0.16/fs/jbd/commit.c
2 ===================================================================
3 --- linux-2.6.16.53-0.16.orig/fs/jbd/commit.c
4 +++ linux-2.6.16.53-0.16/fs/jbd/commit.c
6 #include <linux/pagemap.h>
7 #include <linux/smp_lock.h>
8 #include <linux/jiffies.h>
9 +#include <linux/crc32.h>
12 * Default IO end handler for temporary BJ_IO buffer_heads.
13 @@ -94,19 +95,23 @@ static int inverted_lock(journal_t *jour
17 -/* Done it all: now write the commit record. We should have
19 + * Done it all: now submit the commit record. We should have
20 * cleaned up our previous buffers by now, so if we are in abort
21 * mode we can now just skip the rest of the journal write
24 * Returns 1 if the journal needs to be aborted or 0 on success
26 -static int journal_write_commit_record(journal_t *journal,
27 - transaction_t *commit_transaction)
28 +static int journal_submit_commit_record(journal_t *journal,
29 + transaction_t *commit_transaction,
30 + struct buffer_head **cbh,
33 struct journal_head *descriptor;
34 + struct commit_header *tmp;
35 struct buffer_head *bh;
40 if (is_journal_aborted(journal))
41 @@ -118,21 +123,34 @@ static int journal_write_commit_record(j
43 bh = jh2bh(descriptor);
45 - /* AKPM: buglet - add `i' to tmp! */
46 - for (i = 0; i < bh->b_size; i += 512) {
47 - journal_header_t *tmp = (journal_header_t*)bh->b_data;
48 - tmp->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
49 - tmp->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
50 - tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
51 + tmp = (struct commit_header *)bh->b_data;
52 + tmp->h_magic = cpu_to_be32(JFS_MAGIC_NUMBER);
53 + tmp->h_blocktype = cpu_to_be32(JFS_COMMIT_BLOCK);
54 + tmp->h_sequence = cpu_to_be32(commit_transaction->t_tid);
56 + if (JFS_HAS_COMPAT_FEATURE(journal,
57 + JFS_FEATURE_COMPAT_CHECKSUM)) {
58 + tmp->h_chksum_type = JFS_CRC32_CHKSUM;
59 + tmp->h_chksum_size = JFS_CRC32_CHKSUM_SIZE;
60 + tmp->h_chksum[0] = cpu_to_be32(crc32_sum);
63 - JBUFFER_TRACE(descriptor, "write commit block");
64 + JBUFFER_TRACE(descriptor, "submit commit block");
68 - if (journal->j_flags & JFS_BARRIER) {
69 + set_buffer_uptodate(bh);
70 + bh->b_end_io = journal_end_buffer_io_sync;
72 + if (journal->j_flags & JFS_BARRIER &&
73 + !JFS_HAS_COMPAT_FEATURE(journal,
74 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
76 set_buffer_ordered(bh);
79 - ret = sync_dirty_buffer(bh);
80 + ret = submit_bh(WRITE, bh);
82 /* is it possible for another commit to fail at roughly
83 * the same time as this one? If so, we don't want to
84 * trust the barrier flag in the super, but instead want
85 @@ -153,12 +171,84 @@ static int journal_write_commit_record(j
86 clear_buffer_ordered(bh);
87 set_buffer_uptodate(bh);
89 - ret = sync_dirty_buffer(bh);
90 + ret = submit_bh(WRITE, bh);
92 - put_bh(bh); /* One for getblk() */
93 - journal_put_journal_head(descriptor);
98 - return (ret == -EIO);
100 + * This function along with journal_submit_commit_record
101 + * allows to write the commit record asynchronously.
103 +static int journal_wait_on_commit_record(struct buffer_head *bh)
107 + clear_buffer_dirty(bh);
108 + wait_on_buffer(bh);
110 + if (unlikely(!buffer_uptodate(bh)))
112 + put_bh(bh); /* One for getblk() */
113 + journal_put_journal_head(bh2jh(bh));
119 + * Wait for all submitted IO to complete.
121 +static int journal_wait_on_locked_list(journal_t *journal,
122 + transaction_t *commit_transaction)
125 + struct journal_head *jh;
127 + while (commit_transaction->t_locked_list) {
128 + struct buffer_head *bh;
130 + jh = commit_transaction->t_locked_list->b_tprev;
133 + if (buffer_locked(bh)) {
134 + spin_unlock(&journal->j_list_lock);
135 + wait_on_buffer(bh);
136 + if (unlikely(!buffer_uptodate(bh)))
138 + spin_lock(&journal->j_list_lock);
140 + if (!inverted_lock(journal, bh)) {
142 + spin_lock(&journal->j_list_lock);
145 + if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
146 + __journal_unfile_buffer(jh);
147 + jbd_unlock_bh_state(bh);
148 + journal_remove_journal_head(bh);
151 + jbd_unlock_bh_state(bh);
154 + cond_resched_lock(&journal->j_list_lock);
159 +static inline __u32 jbd_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
161 + struct page *page = bh->b_page;
166 + checksum = crc32_be(crc32_sum,
167 + (void *)(addr + offset_in_page(bh->b_data)),
174 @@ -184,6 +274,8 @@ void journal_commit_transaction(journal_
178 + struct buffer_head *cbh = NULL; /* For transactional checksums */
179 + __u32 crc32_sum = ~0;
182 * First job: lock down the current transaction and wait for
183 @@ -395,37 +487,14 @@ write_out_data:
187 - * Wait for all previously submitted IO to complete.
188 + * Wait for all previously submitted IO to complete if commit
189 + * record is to be written synchronously.
191 - while (commit_transaction->t_locked_list) {
192 - struct buffer_head *bh;
193 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
194 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT))
195 + err = journal_wait_on_locked_list(journal,
196 + commit_transaction);
198 - jh = commit_transaction->t_locked_list->b_tprev;
201 - if (buffer_locked(bh)) {
202 - spin_unlock(&journal->j_list_lock);
203 - wait_on_buffer(bh);
204 - if (unlikely(!buffer_uptodate(bh)))
206 - spin_lock(&journal->j_list_lock);
208 - if (!inverted_lock(journal, bh)) {
210 - spin_lock(&journal->j_list_lock);
213 - if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
214 - __journal_unfile_buffer(jh);
215 - jbd_unlock_bh_state(bh);
216 - journal_remove_journal_head(bh);
219 - jbd_unlock_bh_state(bh);
222 - cond_resched_lock(&journal->j_list_lock);
224 spin_unlock(&journal->j_list_lock);
227 @@ -598,6 +667,16 @@ write_out_data:
229 for (i = 0; i < bufs; i++) {
230 struct buffer_head *bh = wbuf[i];
232 + * Compute checksum.
234 + if (JFS_HAS_COMPAT_FEATURE(journal,
235 + JFS_FEATURE_COMPAT_CHECKSUM)) {
237 + jbd_checksum_data(crc32_sum,
242 clear_buffer_dirty(bh);
243 set_buffer_uptodate(bh);
244 @@ -614,6 +693,23 @@ start_journal_io:
248 + /* Done it all: now write the commit record asynchronously. */
250 + if (JFS_HAS_INCOMPAT_FEATURE(journal,
251 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
252 + err = journal_submit_commit_record(journal, commit_transaction,
255 + __journal_abort_hard(journal);
257 + spin_lock(&journal->j_list_lock);
258 + err = journal_wait_on_locked_list(journal,
259 + commit_transaction);
260 + spin_unlock(&journal->j_list_lock);
262 + __journal_abort_hard(journal);
265 /* Lo and behold: we have just managed to send a transaction to
266 the log. Before we can commit it, wait for the IO so far to
267 complete. Control buffers being written are on the
268 @@ -712,9 +808,15 @@ wait_for_iobuf:
271 jbd_debug(3, "JBD: commit phase 6\n");
273 - if (journal_write_commit_record(journal, commit_transaction))
276 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
277 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
278 + err = journal_submit_commit_record(journal, commit_transaction,
281 + __journal_abort_hard(journal);
283 + err = journal_wait_on_commit_record(cbh);
286 __journal_abort_hard(journal);
287 Index: linux-2.6.16.53-0.16/include/linux/jbd.h
288 ===================================================================
289 --- linux-2.6.16.53-0.16.orig/include/linux/jbd.h
290 +++ linux-2.6.16.53-0.16/include/linux/jbd.h
291 @@ -142,6 +142,29 @@ typedef struct journal_header_s
298 +#define JFS_CRC32_CHKSUM 1
299 +#define JFS_MD5_CHKSUM 2
300 +#define JFS_SHA1_CHKSUM 3
302 +#define JFS_CRC32_CHKSUM_SIZE 4
304 +#define JFS_CHECKSUM_BYTES (32 / sizeof(u32))
306 + * Commit block header for storing transactional checksums:
308 +struct commit_header
311 + __be32 h_blocktype;
313 + unsigned char h_chksum_type;
314 + unsigned char h_chksum_size;
315 + unsigned char h_padding[2];
316 + __be32 h_chksum[JFS_CHECKSUM_BYTES];
320 * The block tag: used to describe a single buffer in the journal
321 @@ -228,12 +251,16 @@ typedef struct journal_superblock_s
322 ((j)->j_format_version >= 2 && \
323 ((j)->j_superblock->s_feature_incompat & cpu_to_be32((mask))))
325 -#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
326 +#define JFS_FEATURE_COMPAT_CHECKSUM 0x00000001
328 +#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
329 +#define JFS_FEATURE_INCOMPAT_ASYNC_COMMIT 0x00000004
331 /* Features known to this kernel version: */
332 -#define JFS_KNOWN_COMPAT_FEATURES 0
333 +#define JFS_KNOWN_COMPAT_FEATURES JFS_FEATURE_COMPAT_CHECKSUM
334 #define JFS_KNOWN_ROCOMPAT_FEATURES 0
335 -#define JFS_KNOWN_INCOMPAT_FEATURES JFS_FEATURE_INCOMPAT_REVOKE
336 +#define JFS_KNOWN_INCOMPAT_FEATURES JFS_FEATURE_INCOMPAT_REVOKE | \
337 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT
341 @@ -1041,6 +1068,8 @@ extern int journal_check_available_fe
342 (journal_t *, unsigned long, unsigned long, unsigned long);
343 extern int journal_set_features
344 (journal_t *, unsigned long, unsigned long, unsigned long);
345 +extern int journal_clear_features
346 + (journal_t *, unsigned long, unsigned long, unsigned long);
347 extern int journal_create (journal_t *);
348 extern int journal_load (journal_t *journal);
349 extern void journal_destroy (journal_t *);
350 Index: linux-2.6.16.53-0.16/fs/jbd/recovery.c
351 ===================================================================
352 --- linux-2.6.16.53-0.16.orig/fs/jbd/recovery.c
353 +++ linux-2.6.16.53-0.16/fs/jbd/recovery.c
355 #include <linux/jbd.h>
356 #include <linux/errno.h>
357 #include <linux/slab.h>
358 +#include <linux/crc32.h>
362 @@ -307,6 +308,37 @@ int journal_skip_recovery(journal_t *jou
367 + * calc_chksums calculates the checksums for the blocks described in the
368 + * descriptor block.
370 +static int calc_chksums(journal_t *journal, struct buffer_head *bh,
371 + unsigned long *next_log_block, __u32 *crc32_sum)
373 + int i, num_blks, err;
375 + struct buffer_head *obh;
377 + num_blks = count_tags(bh, journal->j_blocksize);
378 + /* Calculate checksum of the descriptor block. */
379 + *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
381 + for (i = 0; i < num_blks; i++) {
382 + io_block = (*next_log_block)++;
383 + wrap(journal, *next_log_block);
384 + err = jread(&obh, journal, io_block);
386 + printk(KERN_ERR "JBD: IO error %d recovering block "
387 + "%u in log\n", err, io_block);
390 + *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
397 static int do_one_pass(journal_t *journal,
398 struct recovery_info *info, enum passtype pass)
400 @@ -318,6 +350,7 @@ static int do_one_pass(journal_t *journa
401 struct buffer_head * bh;
402 unsigned int sequence;
404 + __u32 crc32_sum = ~0; /* Transactional Checksums */
406 /* Precompute the maximum metadata descriptors in a descriptor block */
407 int MAX_BLOCKS_PER_DESC;
408 @@ -409,9 +442,24 @@ static int do_one_pass(journal_t *journa
410 case JFS_DESCRIPTOR_BLOCK:
411 /* If it is a valid descriptor block, replay it
412 - * in pass REPLAY; otherwise, just skip over the
413 - * blocks it describes. */
414 + * in pass REPLAY; if journal_checksums enabled, then
415 + * calculate checksums in PASS_SCAN, otherwise,
416 + * just skip over the blocks it describes. */
417 if (pass != PASS_REPLAY) {
418 + if (pass == PASS_SCAN &&
419 + JFS_HAS_COMPAT_FEATURE(journal,
420 + JFS_FEATURE_COMPAT_CHECKSUM) &&
421 + !info->end_transaction) {
422 + if (calc_chksums(journal, bh,
433 count_tags(bh, journal->j_blocksize);
434 wrap(journal, next_log_block);
435 @@ -506,9 +554,97 @@ static int do_one_pass(journal_t *journa
438 case JFS_COMMIT_BLOCK:
439 - /* Found an expected commit block: not much to
440 - * do other than move on to the next sequence
441 + /* How to differentiate between interrupted commit
442 + * and journal corruption ?
444 + * {nth transaction}
445 + * Checksum Verification Failed
447 + * ____________________
449 + * async_commit sync_commit
451 + * | GO TO NEXT "Journal Corruption"
454 + * {(n+1)th transanction}
456 + * _______|______________
458 + * Commit block found Commit block not found
460 + * "Journal Corruption" |
461 + * _____________|__________
463 + * nth trans corrupt OR nth trans
464 + * and (n+1)th interrupted interrupted
465 + * before commit block
466 + * could reach the disk.
467 + * (Cannot find the difference in above
468 + * mentioned conditions. Hence assume
469 + * "Interrupted Commit".)
472 + /* Found an expected commit block: if checksums
473 + * are present verify them in PASS_SCAN; else not
474 + * much to do other than move on to the next sequence
476 + if (pass == PASS_SCAN &&
477 + JFS_HAS_COMPAT_FEATURE(journal,
478 + JFS_FEATURE_COMPAT_CHECKSUM)) {
479 + int chksum_err, chksum_seen;
480 + struct commit_header *cbh =
481 + (struct commit_header *)bh->b_data;
482 + unsigned found_chksum =
483 + be32_to_cpu(cbh->h_chksum[0]);
485 + chksum_err = chksum_seen = 0;
487 + if (info->end_transaction) {
488 + printk(KERN_ERR "JBD: Transaction %u "
489 + "found to be corrupt.\n",
490 + next_commit_ID - 1);
495 + if (crc32_sum == found_chksum &&
496 + cbh->h_chksum_type == JFS_CRC32_CHKSUM &&
497 + cbh->h_chksum_size ==
498 + JFS_CRC32_CHKSUM_SIZE) {
500 + } else if (!(cbh->h_chksum_type == 0 &&
501 + cbh->h_chksum_size == 0 &&
502 + found_chksum == 0 &&
505 + * If fs is mounted using an old kernel and then
506 + * kernel with journal_chksum is used then we
507 + * get a situation where the journal flag has
508 + * checksum flag set but checksums are not
509 + * present i.e chksum = 0, in the individual
511 + * Hence to avoid checksum failures, in this
512 + * situation, this extra check is added.
518 + info->end_transaction = next_commit_ID;
520 + if (!JFS_HAS_COMPAT_FEATURE(journal,
521 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){
523 + "JBD: Transaction %u "
524 + "found to be corrupt.\n",
535 @@ -543,9 +679,10 @@ static int do_one_pass(journal_t *journa
536 * transaction marks the end of the valid log.
539 - if (pass == PASS_SCAN)
540 - info->end_transaction = next_commit_ID;
542 + if (pass == PASS_SCAN) {
543 + if (!info->end_transaction)
544 + info->end_transaction = next_commit_ID;
546 /* It's really bad news if different passes end up at
547 * different places (but possible due to IO errors). */
548 if (info->end_transaction != next_commit_ID) {
549 Index: linux-2.6.16.53-0.16/fs/jbd/journal.c
550 ===================================================================
551 --- linux-2.6.16.53-0.16.orig/fs/jbd/journal.c
552 +++ linux-2.6.16.53-0.16/fs/jbd/journal.c
553 @@ -64,6 +64,7 @@ EXPORT_SYMBOL(journal_update_format);
554 EXPORT_SYMBOL(journal_check_used_features);
555 EXPORT_SYMBOL(journal_check_available_features);
556 EXPORT_SYMBOL(journal_set_features);
557 +EXPORT_SYMBOL(journal_clear_features);
558 EXPORT_SYMBOL(journal_create);
559 EXPORT_SYMBOL(journal_load);
560 EXPORT_SYMBOL(journal_destroy);
561 @@ -1565,6 +1566,33 @@ int journal_set_features (journal_t *jou
566 + * int journal_clear_features () - Clear a given journal feature in the superblock
567 + * @journal: Journal to act on.
568 + * @compat: bitmask of compatible features
569 + * @ro: bitmask of features that force read-only mount
570 + * @incompat: bitmask of incompatible features
572 + * Clear a given journal feature as present on the
573 + * superblock. Returns true if the requested features could be reset.
576 +int journal_clear_features (journal_t *journal, unsigned long compat,
577 + unsigned long ro, unsigned long incompat)
579 + journal_superblock_t *sb;
581 + jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
582 + compat, ro, incompat);
584 + sb = journal->j_superblock;
586 + sb->s_feature_compat &= ~cpu_to_be32(compat);
587 + sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
588 + sb->s_feature_incompat &= ~cpu_to_be32(incompat);
594 * int journal_update_format () - Update on-disk journal structure.
595 Index: linux-2.6.16.53-0.16/fs/Kconfig
596 ===================================================================
597 --- linux-2.6.16.53-0.16.orig/fs/Kconfig
598 +++ linux-2.6.16.53-0.16/fs/Kconfig
599 @@ -140,6 +140,7 @@ config EXT3_FS_SECURITY
605 This is a generic journaling layer for block devices. It is
606 currently used by the ext3 and OCFS2 file systems, but it could
607 Index: linux-2.6.16.53-0.16/Documentation/filesystems/ext3.txt
608 ===================================================================
609 --- linux-2.6.16.53-0.16.orig/Documentation/filesystems/ext3.txt
610 +++ linux-2.6.16.53-0.16/Documentation/filesystems/ext3.txt
611 @@ -14,6 +14,16 @@ Options
612 When mounting an ext3 filesystem, the following option are accepted:
615 +journal_checksum Enable checksumming of the journal transactions.
616 + This will allow the recovery code in e2fsck and the
617 + kernel to detect corruption in the kernel. It is a
618 + compatible change and will be ignored by older kernels.
620 +journal_async_commit Commit block can be written to disk without waiting
621 + for descriptor blocks. If enabled older kernels cannot
622 + mount the device. This will enable 'journal_checksum'
625 journal=update Update the ext3 file system's journal to the current