1 Index: linux-2.6.18-8.1.8/fs/jbd/commit.c
2 ===================================================================
3 --- linux-2.6.18-8.1.8.orig/fs/jbd/commit.c
4 +++ linux-2.6.18-8.1.8/fs/jbd/commit.c
7 #include <linux/pagemap.h>
8 #include <linux/smp_lock.h>
9 +#include <linux/crc32.h>
12 * Default IO end handler for temporary BJ_IO buffer_heads.
13 @@ -93,19 +94,23 @@ 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 @@ -117,21 +122,34 @@ 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 @@ -152,14 +170,72 @@ 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 void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
162 @@ -293,6 +369,8 @@ void journal_commit_transaction(journal_
166 + struct buffer_head *cbh = NULL; /* For transactional checksums */
167 + __u32 crc32_sum = ~0;
170 * First job: lock down the current transaction and wait for
171 @@ -428,38 +506,14 @@ void journal_commit_transaction(journal_
172 journal_submit_data_buffers(journal, commit_transaction);
175 - * Wait for all previously submitted IO to complete.
176 + * Wait for all previously submitted IO to complete if commit
177 + * record is to be written synchronously.
179 spin_lock(&journal->j_list_lock);
180 - while (commit_transaction->t_locked_list) {
181 - struct buffer_head *bh;
183 - jh = commit_transaction->t_locked_list->b_tprev;
186 - if (buffer_locked(bh)) {
187 - spin_unlock(&journal->j_list_lock);
188 - wait_on_buffer(bh);
189 - if (unlikely(!buffer_uptodate(bh)))
191 - spin_lock(&journal->j_list_lock);
193 - if (!inverted_lock(journal, bh)) {
195 - spin_lock(&journal->j_list_lock);
198 - if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
199 - __journal_unfile_buffer(jh);
200 - jbd_unlock_bh_state(bh);
201 - journal_remove_journal_head(bh);
204 - jbd_unlock_bh_state(bh);
207 - cond_resched_lock(&journal->j_list_lock);
209 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
210 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT))
211 + err = journal_wait_on_locked_list(journal,
212 + commit_transaction);
213 spin_unlock(&journal->j_list_lock);
216 @@ -627,6 +681,16 @@ void journal_commit_transaction(journal_
218 for (i = 0; i < bufs; i++) {
219 struct buffer_head *bh = wbuf[i];
221 + * Compute checksum.
223 + if (JFS_HAS_COMPAT_FEATURE(journal,
224 + JFS_FEATURE_COMPAT_CHECKSUM)) {
225 + crc32_sum = crc32_be(crc32_sum,
226 + (void *)bh->b_data,
231 clear_buffer_dirty(bh);
232 set_buffer_uptodate(bh);
233 @@ -642,6 +706,23 @@ start_journal_io:
237 + /* Done it all: now write the commit record asynchronously. */
239 + if (JFS_HAS_INCOMPAT_FEATURE(journal,
240 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
241 + err = journal_submit_commit_record(journal, commit_transaction,
244 + __journal_abort_hard(journal);
246 + spin_lock(&journal->j_list_lock);
247 + err = journal_wait_on_locked_list(journal,
248 + commit_transaction);
249 + spin_unlock(&journal->j_list_lock);
251 + __journal_abort_hard(journal);
254 /* Lo and behold: we have just managed to send a transaction to
255 the log. Before we can commit it, wait for the IO so far to
256 complete. Control buffers being written are on the
257 @@ -740,9 +821,15 @@ wait_for_iobuf:
260 jbd_debug(3, "JBD: commit phase 6\n");
262 - if (journal_write_commit_record(journal, commit_transaction))
265 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
266 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
267 + err = journal_submit_commit_record(journal, commit_transaction,
270 + __journal_abort_hard(journal);
272 + err = journal_wait_on_commit_record(cbh);
275 __journal_abort_hard(journal);
276 Index: linux-2.6.18-8.1.8/include/linux/jbd.h
277 ===================================================================
278 --- linux-2.6.18-8.1.8.orig/include/linux/jbd.h
279 +++ linux-2.6.18-8.1.8/include/linux/jbd.h
280 @@ -148,6 +148,29 @@ typedef struct journal_header_s
287 +#define JFS_CRC32_CHKSUM 1
288 +#define JFS_MD5_CHKSUM 2
289 +#define JFS_SHA1_CHKSUM 3
291 +#define JFS_CRC32_CHKSUM_SIZE 4
293 +#define JFS_CHECKSUM_BYTES (32 / sizeof(u32))
295 + * Commit block header for storing transactional checksums:
297 +struct commit_header
300 + __be32 h_blocktype;
302 + unsigned char h_chksum_type;
303 + unsigned char h_chksum_size;
304 + unsigned char h_padding[2];
305 + __be32 h_chksum[JFS_CHECKSUM_BYTES];
309 * The block tag: used to describe a single buffer in the journal
310 @@ -234,12 +257,16 @@ typedef struct journal_superblock_s
311 ((j)->j_format_version >= 2 && \
312 ((j)->j_superblock->s_feature_incompat & cpu_to_be32((mask))))
314 -#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
315 +#define JFS_FEATURE_COMPAT_CHECKSUM 0x00000001
317 +#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
318 +#define JFS_FEATURE_INCOMPAT_ASYNC_COMMIT 0x00000004
320 /* Features known to this kernel version: */
321 -#define JFS_KNOWN_COMPAT_FEATURES 0
322 +#define JFS_KNOWN_COMPAT_FEATURES JFS_FEATURE_COMPAT_CHECKSUM
323 #define JFS_KNOWN_ROCOMPAT_FEATURES 0
324 -#define JFS_KNOWN_INCOMPAT_FEATURES JFS_FEATURE_INCOMPAT_REVOKE
325 +#define JFS_KNOWN_INCOMPAT_FEATURES JFS_FEATURE_INCOMPAT_REVOKE | \
326 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT
330 @@ -967,6 +994,8 @@ extern int journal_check_available_fe
331 (journal_t *, unsigned long, unsigned long, unsigned long);
332 extern int journal_set_features
333 (journal_t *, unsigned long, unsigned long, unsigned long);
334 +extern int journal_clear_features
335 + (journal_t *, unsigned long, unsigned long, unsigned long);
336 extern int journal_create (journal_t *);
337 extern int journal_load (journal_t *journal);
338 extern void journal_destroy (journal_t *);
339 Index: linux-2.6.18-8.1.8/fs/jbd/recovery.c
340 ===================================================================
341 --- linux-2.6.18-8.1.8.orig/fs/jbd/recovery.c
342 +++ linux-2.6.18-8.1.8/fs/jbd/recovery.c
344 #include <linux/jbd.h>
345 #include <linux/errno.h>
346 #include <linux/slab.h>
347 +#include <linux/crc32.h>
351 @@ -307,6 +308,37 @@ int journal_skip_recovery(journal_t *jou
356 + * calc_chksums calculates the checksums for the blocks described in the
357 + * descriptor block.
359 +static int calc_chksums(journal_t *journal, struct buffer_head *bh,
360 + unsigned long *next_log_block, __u32 *crc32_sum)
362 + int i, num_blks, err;
364 + struct buffer_head *obh;
366 + num_blks = count_tags(bh, journal->j_blocksize);
367 + /* Calculate checksum of the descriptor block. */
368 + *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
370 + for (i = 0; i < num_blks; i++) {
371 + io_block = (*next_log_block)++;
372 + wrap(journal, *next_log_block);
373 + err = jread(&obh, journal, io_block);
375 + printk(KERN_ERR "JBD: IO error %d recovering block "
376 + "%u in log\n", err, io_block);
379 + *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
386 static int do_one_pass(journal_t *journal,
387 struct recovery_info *info, enum passtype pass)
389 @@ -318,6 +350,7 @@ static int do_one_pass(journal_t *journa
390 struct buffer_head * bh;
391 unsigned int sequence;
393 + __u32 crc32_sum = ~0; /* Transactional Checksums */
395 /* Precompute the maximum metadata descriptors in a descriptor block */
396 int MAX_BLOCKS_PER_DESC;
397 @@ -409,9 +442,24 @@ static int do_one_pass(journal_t *journa
399 case JFS_DESCRIPTOR_BLOCK:
400 /* If it is a valid descriptor block, replay it
401 - * in pass REPLAY; otherwise, just skip over the
402 - * blocks it describes. */
403 + * in pass REPLAY; if journal_checksums enabled, then
404 + * calculate checksums in PASS_SCAN, otherwise,
405 + * just skip over the blocks it describes. */
406 if (pass != PASS_REPLAY) {
407 + if (pass == PASS_SCAN &&
408 + JFS_HAS_COMPAT_FEATURE(journal,
409 + JFS_FEATURE_COMPAT_CHECKSUM) &&
410 + !info->end_transaction) {
411 + if (calc_chksums(journal, bh,
422 count_tags(bh, journal->j_blocksize);
423 wrap(journal, next_log_block);
424 @@ -506,9 +554,97 @@ static int do_one_pass(journal_t *journa
427 case JFS_COMMIT_BLOCK:
428 - /* Found an expected commit block: not much to
429 - * do other than move on to the next sequence
430 + /* How to differentiate between interrupted commit
431 + * and journal corruption ?
433 + * {nth transaction}
434 + * Checksum Verification Failed
436 + * ____________________
438 + * async_commit sync_commit
440 + * | GO TO NEXT "Journal Corruption"
443 + * {(n+1)th transanction}
445 + * _______|______________
447 + * Commit block found Commit block not found
449 + * "Journal Corruption" |
450 + * _____________|__________
452 + * nth trans corrupt OR nth trans
453 + * and (n+1)th interrupted interrupted
454 + * before commit block
455 + * could reach the disk.
456 + * (Cannot find the difference in above
457 + * mentioned conditions. Hence assume
458 + * "Interrupted Commit".)
461 + /* Found an expected commit block: if checksums
462 + * are present verify them in PASS_SCAN; else not
463 + * much to do other than move on to the next sequence
465 + if (pass == PASS_SCAN &&
466 + JFS_HAS_COMPAT_FEATURE(journal,
467 + JFS_FEATURE_COMPAT_CHECKSUM)) {
468 + int chksum_err, chksum_seen;
469 + struct commit_header *cbh =
470 + (struct commit_header *)bh->b_data;
471 + unsigned found_chksum =
472 + be32_to_cpu(cbh->h_chksum[0]);
474 + chksum_err = chksum_seen = 0;
476 + if (info->end_transaction) {
477 + printk(KERN_ERR "JBD: Transaction %u "
478 + "found to be corrupt.\n",
479 + next_commit_ID - 1);
484 + if (crc32_sum == found_chksum &&
485 + cbh->h_chksum_type == JFS_CRC32_CHKSUM &&
486 + cbh->h_chksum_size ==
487 + JFS_CRC32_CHKSUM_SIZE) {
489 + } else if (!(cbh->h_chksum_type == 0 &&
490 + cbh->h_chksum_size == 0 &&
491 + found_chksum == 0 &&
494 + * If fs is mounted using an old kernel and then
495 + * kernel with journal_chksum is used then we
496 + * get a situation where the journal flag has
497 + * checksum flag set but checksums are not
498 + * present i.e chksum = 0, in the individual
500 + * Hence to avoid checksum failures, in this
501 + * situation, this extra check is added.
507 + info->end_transaction = next_commit_ID;
509 + if (!JFS_HAS_COMPAT_FEATURE(journal,
510 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){
512 + "JBD: Transaction %u "
513 + "found to be corrupt.\n",
524 @@ -544,9 +680,10 @@ static int do_one_pass(journal_t *journa
525 * transaction marks the end of the valid log.
528 - if (pass == PASS_SCAN)
529 - info->end_transaction = next_commit_ID;
531 + if (pass == PASS_SCAN) {
532 + if (!info->end_transaction)
533 + info->end_transaction = next_commit_ID;
535 /* It's really bad news if different passes end up at
536 * different places (but possible due to IO errors). */
537 if (info->end_transaction != next_commit_ID) {
538 Index: linux-2.6.18-8.1.8/fs/jbd/journal.c
539 ===================================================================
540 --- linux-2.6.18-8.1.8.orig/fs/jbd/journal.c
541 +++ linux-2.6.18-8.1.8/fs/jbd/journal.c
542 @@ -66,6 +66,7 @@ EXPORT_SYMBOL(journal_update_format);
543 EXPORT_SYMBOL(journal_check_used_features);
544 EXPORT_SYMBOL(journal_check_available_features);
545 EXPORT_SYMBOL(journal_set_features);
546 +EXPORT_SYMBOL(journal_clear_features);
547 EXPORT_SYMBOL(journal_create);
548 EXPORT_SYMBOL(journal_load);
549 EXPORT_SYMBOL(journal_destroy);
550 @@ -1271,6 +1272,33 @@ int journal_set_features (journal_t *jou
555 + * int journal_clear_features () - Clear a given journal feature in the superblock
556 + * @journal: Journal to act on.
557 + * @compat: bitmask of compatible features
558 + * @ro: bitmask of features that force read-only mount
559 + * @incompat: bitmask of incompatible features
561 + * Clear a given journal feature as present on the
562 + * superblock. Returns true if the requested features could be reset.
565 +int journal_clear_features (journal_t *journal, unsigned long compat,
566 + unsigned long ro, unsigned long incompat)
568 + journal_superblock_t *sb;
570 + jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
571 + compat, ro, incompat);
573 + sb = journal->j_superblock;
575 + sb->s_feature_compat &= ~cpu_to_be32(compat);
576 + sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
577 + sb->s_feature_incompat &= ~cpu_to_be32(incompat);
583 * int journal_update_format () - Update on-disk journal structure.
584 Index: linux-2.6.18-8.1.8/fs/Kconfig
585 ===================================================================
586 --- linux-2.6.18-8.1.8.orig/fs/Kconfig
587 +++ linux-2.6.18-8.1.8/fs/Kconfig
588 @@ -140,6 +140,7 @@ config EXT3_FS_SECURITY
594 This is a generic journaling layer for block devices. It is
595 currently used by the ext3 and OCFS2 file systems, but it could
596 Index: linux-2.6.18-8.1.8/Documentation/filesystems/ext3.txt
597 ===================================================================
598 --- linux-2.6.18-8.1.8.orig/Documentation/filesystems/ext3.txt
599 +++ linux-2.6.18-8.1.8/Documentation/filesystems/ext3.txt
600 @@ -14,6 +14,16 @@ Options
601 When mounting an ext3 filesystem, the following option are accepted:
604 +journal_checksum Enable checksumming of the journal transactions.
605 + This will allow the recovery code in e2fsck and the
606 + kernel to detect corruption in the kernel. It is a
607 + compatible change and will be ignored by older kernels.
609 +journal_async_commit Commit block can be written to disk without waiting
610 + for descriptor blocks. If enabled older kernels cannot
611 + mount the device. This will enable 'journal_checksum'
614 journal=update Update the ext3 file system's journal to the current