1 Index: linux-2.6.16.46-0.14/fs/jbd/commit.c
2 ===================================================================
3 --- linux-2.6.16.46-0.14.orig/fs/jbd/commit.c
4 +++ linux-2.6.16.46-0.14/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 @@ -153,15 +171,74 @@ 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);
161 * journal_commit_transaction
163 * The primary function for committing a transaction to the log. This
164 @@ -184,6 +261,8 @@ void journal_commit_transaction(journal_
168 + struct buffer_head *cbh = NULL; /* For transactional checksums */
169 + __u32 crc32_sum = ~0;
172 * First job: lock down the current transaction and wait for
173 @@ -395,37 +474,14 @@ write_out_data:
177 - * Wait for all previously submitted IO to complete.
178 + * Wait for all previously submitted IO to complete if commit
179 + * record is to be written synchronously.
181 - while (commit_transaction->t_locked_list) {
182 - struct buffer_head *bh;
183 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
184 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT))
185 + err = journal_wait_on_locked_list(journal,
186 + commit_transaction);
188 - jh = commit_transaction->t_locked_list->b_tprev;
191 - if (buffer_locked(bh)) {
192 - spin_unlock(&journal->j_list_lock);
193 - wait_on_buffer(bh);
194 - if (unlikely(!buffer_uptodate(bh)))
196 - spin_lock(&journal->j_list_lock);
198 - if (!inverted_lock(journal, bh)) {
200 - spin_lock(&journal->j_list_lock);
203 - if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
204 - __journal_unfile_buffer(jh);
205 - jbd_unlock_bh_state(bh);
206 - journal_remove_journal_head(bh);
209 - jbd_unlock_bh_state(bh);
212 - cond_resched_lock(&journal->j_list_lock);
214 spin_unlock(&journal->j_list_lock);
217 @@ -598,6 +654,16 @@ write_out_data:
219 for (i = 0; i < bufs; i++) {
220 struct buffer_head *bh = wbuf[i];
222 + * Compute checksum.
224 + if (JFS_HAS_COMPAT_FEATURE(journal,
225 + JFS_FEATURE_COMPAT_CHECKSUM)) {
226 + crc32_sum = crc32_be(crc32_sum,
227 + (void *)bh->b_data,
232 clear_buffer_dirty(bh);
233 set_buffer_uptodate(bh);
234 @@ -614,6 +680,23 @@ start_journal_io:
238 + /* Done it all: now write the commit record asynchronously. */
240 + if (JFS_HAS_INCOMPAT_FEATURE(journal,
241 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
242 + err = journal_submit_commit_record(journal, commit_transaction,
245 + __journal_abort_hard(journal);
247 + spin_lock(&journal->j_list_lock);
248 + err = journal_wait_on_locked_list(journal,
249 + commit_transaction);
250 + spin_unlock(&journal->j_list_lock);
252 + __journal_abort_hard(journal);
255 /* Lo and behold: we have just managed to send a transaction to
256 the log. Before we can commit it, wait for the IO so far to
257 complete. Control buffers being written are on the
258 @@ -712,9 +795,15 @@ wait_for_iobuf:
261 jbd_debug(3, "JBD: commit phase 6\n");
263 - if (journal_write_commit_record(journal, commit_transaction))
266 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
267 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
268 + err = journal_submit_commit_record(journal, commit_transaction,
271 + __journal_abort_hard(journal);
273 + err = journal_wait_on_commit_record(cbh);
276 __journal_abort_hard(journal);
277 Index: linux-2.6.16.46-0.14/include/linux/jbd.h
278 ===================================================================
279 --- linux-2.6.16.46-0.14.orig/include/linux/jbd.h
280 +++ linux-2.6.16.46-0.14/include/linux/jbd.h
281 @@ -142,6 +142,29 @@ typedef struct journal_header_s
288 +#define JFS_CRC32_CHKSUM 1
289 +#define JFS_MD5_CHKSUM 2
290 +#define JFS_SHA1_CHKSUM 3
292 +#define JFS_CRC32_CHKSUM_SIZE 4
294 +#define JFS_CHECKSUM_BYTES (32 / sizeof(u32))
296 + * Commit block header for storing transactional checksums:
298 +struct commit_header
301 + __be32 h_blocktype;
303 + unsigned char h_chksum_type;
304 + unsigned char h_chksum_size;
305 + unsigned char h_padding[2];
306 + __be32 h_chksum[JFS_CHECKSUM_BYTES];
310 * The block tag: used to describe a single buffer in the journal
311 @@ -228,12 +251,16 @@ typedef struct journal_superblock_s
312 ((j)->j_format_version >= 2 && \
313 ((j)->j_superblock->s_feature_incompat & cpu_to_be32((mask))))
315 -#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
316 +#define JFS_FEATURE_COMPAT_CHECKSUM 0x00000001
318 +#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
319 +#define JFS_FEATURE_INCOMPAT_ASYNC_COMMIT 0x00000004
321 /* Features known to this kernel version: */
322 -#define JFS_KNOWN_COMPAT_FEATURES 0
323 +#define JFS_KNOWN_COMPAT_FEATURES JFS_FEATURE_COMPAT_CHECKSUM
324 #define JFS_KNOWN_ROCOMPAT_FEATURES 0
325 -#define JFS_KNOWN_INCOMPAT_FEATURES JFS_FEATURE_INCOMPAT_REVOKE
326 +#define JFS_KNOWN_INCOMPAT_FEATURES JFS_FEATURE_INCOMPAT_REVOKE | \
327 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT
331 @@ -1041,6 +1068,8 @@ extern int journal_check_available_fe
332 (journal_t *, unsigned long, unsigned long, unsigned long);
333 extern int journal_set_features
334 (journal_t *, unsigned long, unsigned long, unsigned long);
335 +extern int journal_clear_features
336 + (journal_t *, unsigned long, unsigned long, unsigned long);
337 extern int journal_create (journal_t *);
338 extern int journal_load (journal_t *journal);
339 extern void journal_destroy (journal_t *);
340 Index: linux-2.6.16.46-0.14/fs/jbd/recovery.c
341 ===================================================================
342 --- linux-2.6.16.46-0.14.orig/fs/jbd/recovery.c
343 +++ linux-2.6.16.46-0.14/fs/jbd/recovery.c
345 #include <linux/jbd.h>
346 #include <linux/errno.h>
347 #include <linux/slab.h>
348 +#include <linux/crc32.h>
352 @@ -307,6 +308,37 @@ int journal_skip_recovery(journal_t *jou
357 + * calc_chksums calculates the checksums for the blocks described in the
358 + * descriptor block.
360 +static int calc_chksums(journal_t *journal, struct buffer_head *bh,
361 + unsigned long *next_log_block, __u32 *crc32_sum)
363 + int i, num_blks, err;
365 + struct buffer_head *obh;
367 + num_blks = count_tags(bh, journal->j_blocksize);
368 + /* Calculate checksum of the descriptor block. */
369 + *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
371 + for (i = 0; i < num_blks; i++) {
372 + io_block = (*next_log_block)++;
373 + wrap(journal, *next_log_block);
374 + err = jread(&obh, journal, io_block);
376 + printk(KERN_ERR "JBD: IO error %d recovering block "
377 + "%u in log\n", err, io_block);
380 + *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
387 static int do_one_pass(journal_t *journal,
388 struct recovery_info *info, enum passtype pass)
390 @@ -318,6 +350,7 @@ static int do_one_pass(journal_t *journa
391 struct buffer_head * bh;
392 unsigned int sequence;
394 + __u32 crc32_sum = ~0; /* Transactional Checksums */
396 /* Precompute the maximum metadata descriptors in a descriptor block */
397 int MAX_BLOCKS_PER_DESC;
398 @@ -409,9 +442,24 @@ static int do_one_pass(journal_t *journa
400 case JFS_DESCRIPTOR_BLOCK:
401 /* If it is a valid descriptor block, replay it
402 - * in pass REPLAY; otherwise, just skip over the
403 - * blocks it describes. */
404 + * in pass REPLAY; if journal_checksums enabled, then
405 + * calculate checksums in PASS_SCAN, otherwise,
406 + * just skip over the blocks it describes. */
407 if (pass != PASS_REPLAY) {
408 + if (pass == PASS_SCAN &&
409 + JFS_HAS_COMPAT_FEATURE(journal,
410 + JFS_FEATURE_COMPAT_CHECKSUM) &&
411 + !info->end_transaction) {
412 + if (calc_chksums(journal, bh,
423 count_tags(bh, journal->j_blocksize);
424 wrap(journal, next_log_block);
425 @@ -506,9 +554,97 @@ static int do_one_pass(journal_t *journa
428 case JFS_COMMIT_BLOCK:
429 - /* Found an expected commit block: not much to
430 - * do other than move on to the next sequence
431 + /* How to differentiate between interrupted commit
432 + * and journal corruption ?
434 + * {nth transaction}
435 + * Checksum Verification Failed
437 + * ____________________
439 + * async_commit sync_commit
441 + * | GO TO NEXT "Journal Corruption"
444 + * {(n+1)th transanction}
446 + * _______|______________
448 + * Commit block found Commit block not found
450 + * "Journal Corruption" |
451 + * _____________|__________
453 + * nth trans corrupt OR nth trans
454 + * and (n+1)th interrupted interrupted
455 + * before commit block
456 + * could reach the disk.
457 + * (Cannot find the difference in above
458 + * mentioned conditions. Hence assume
459 + * "Interrupted Commit".)
462 + /* Found an expected commit block: if checksums
463 + * are present verify them in PASS_SCAN; else not
464 + * much to do other than move on to the next sequence
466 + if (pass == PASS_SCAN &&
467 + JFS_HAS_COMPAT_FEATURE(journal,
468 + JFS_FEATURE_COMPAT_CHECKSUM)) {
469 + int chksum_err, chksum_seen;
470 + struct commit_header *cbh =
471 + (struct commit_header *)bh->b_data;
472 + unsigned found_chksum =
473 + be32_to_cpu(cbh->h_chksum[0]);
475 + chksum_err = chksum_seen = 0;
477 + if (info->end_transaction) {
478 + printk(KERN_ERR "JBD: Transaction %u "
479 + "found to be corrupt.\n",
480 + next_commit_ID - 1);
485 + if (crc32_sum == found_chksum &&
486 + cbh->h_chksum_type == JFS_CRC32_CHKSUM &&
487 + cbh->h_chksum_size ==
488 + JFS_CRC32_CHKSUM_SIZE) {
490 + } else if (!(cbh->h_chksum_type == 0 &&
491 + cbh->h_chksum_size == 0 &&
492 + found_chksum == 0 &&
495 + * If fs is mounted using an old kernel and then
496 + * kernel with journal_chksum is used then we
497 + * get a situation where the journal flag has
498 + * checksum flag set but checksums are not
499 + * present i.e chksum = 0, in the individual
501 + * Hence to avoid checksum failures, in this
502 + * situation, this extra check is added.
508 + info->end_transaction = next_commit_ID;
510 + if (!JFS_HAS_COMPAT_FEATURE(journal,
511 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){
513 + "JBD: Transaction %u "
514 + "found to be corrupt.\n",
525 @@ -543,9 +679,10 @@ static int do_one_pass(journal_t *journa
526 * transaction marks the end of the valid log.
529 - if (pass == PASS_SCAN)
530 - info->end_transaction = next_commit_ID;
532 + if (pass == PASS_SCAN) {
533 + if (!info->end_transaction)
534 + info->end_transaction = next_commit_ID;
536 /* It's really bad news if different passes end up at
537 * different places (but possible due to IO errors). */
538 if (info->end_transaction != next_commit_ID) {
539 Index: linux-2.6.16.46-0.14/fs/jbd/journal.c
540 ===================================================================
541 --- linux-2.6.16.46-0.14.orig/fs/jbd/journal.c
542 +++ linux-2.6.16.46-0.14/fs/jbd/journal.c
543 @@ -64,6 +64,7 @@ EXPORT_SYMBOL(journal_update_format);
544 EXPORT_SYMBOL(journal_check_used_features);
545 EXPORT_SYMBOL(journal_check_available_features);
546 EXPORT_SYMBOL(journal_set_features);
547 +EXPORT_SYMBOL(journal_clear_features);
548 EXPORT_SYMBOL(journal_create);
549 EXPORT_SYMBOL(journal_load);
550 EXPORT_SYMBOL(journal_destroy);
551 @@ -1565,6 +1566,33 @@ int journal_set_features (journal_t *jou
556 + * int journal_clear_features () - Clear a given journal feature in the superblock
557 + * @journal: Journal to act on.
558 + * @compat: bitmask of compatible features
559 + * @ro: bitmask of features that force read-only mount
560 + * @incompat: bitmask of incompatible features
562 + * Clear a given journal feature as present on the
563 + * superblock. Returns true if the requested features could be reset.
566 +int journal_clear_features (journal_t *journal, unsigned long compat,
567 + unsigned long ro, unsigned long incompat)
569 + journal_superblock_t *sb;
571 + jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
572 + compat, ro, incompat);
574 + sb = journal->j_superblock;
576 + sb->s_feature_compat &= ~cpu_to_be32(compat);
577 + sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
578 + sb->s_feature_incompat &= ~cpu_to_be32(incompat);
584 * int journal_update_format () - Update on-disk journal structure.
585 Index: linux-2.6.16.46-0.14/fs/Kconfig
586 ===================================================================
587 --- linux-2.6.16.46-0.14.orig/fs/Kconfig
588 +++ linux-2.6.16.46-0.14/fs/Kconfig
589 @@ -140,6 +140,7 @@ config EXT3_FS_SECURITY
595 This is a generic journaling layer for block devices. It is
596 currently used by the ext3 and OCFS2 file systems, but it could
597 Index: linux-2.6.16.46-0.14/Documentation/filesystems/ext3.txt
598 ===================================================================
599 --- linux-2.6.16.46-0.14.orig/Documentation/filesystems/ext3.txt
600 +++ linux-2.6.16.46-0.14/Documentation/filesystems/ext3.txt
601 @@ -14,6 +14,16 @@ Options
602 When mounting an ext3 filesystem, the following option are accepted:
605 +journal_checksum Enable checksumming of the journal transactions.
606 + This will allow the recovery code in e2fsck and the
607 + kernel to detect corruption in the kernel. It is a
608 + compatible change and will be ignored by older kernels.
610 +journal_async_commit Commit block can be written to disk without waiting
611 + for descriptor blocks. If enabled older kernels cannot
612 + mount the device. This will enable 'journal_checksum'
615 journal=update Update the ext3 file system's journal to the current