1 Index: linux-2.6.18.8/fs/jbd/commit.c
2 ===================================================================
3 --- linux-2.6.18.8.orig/fs/jbd/commit.c
4 +++ linux-2.6.18.8/fs/jbd/commit.c
7 #include <linux/pagemap.h>
8 #include <linux/smp_lock.h>
9 +#include <linux/crc32.h>
13 @@ -95,19 +96,23 @@ 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 @@ -119,21 +124,35 @@ 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");
69 - if (journal->j_flags & JFS_BARRIER) {
70 + set_buffer_uptodate(bh);
71 + bh->b_end_io = journal_end_buffer_io_sync;
73 + if (journal->j_flags & JFS_BARRIER &&
74 + !JFS_HAS_INCOMPAT_FEATURE(journal,
75 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
77 set_buffer_ordered(bh);
80 - ret = sync_dirty_buffer(bh);
81 + ret = submit_bh(WRITE, bh);
83 /* is it possible for another commit to fail at roughly
84 * the same time as this one? If so, we don't want to
85 * trust the barrier flag in the super, but instead want
86 @@ -154,12 +173,70 @@ static int journal_write_commit_record(j
87 clear_buffer_ordered(bh);
88 set_buffer_uptodate(bh);
90 - ret = sync_dirty_buffer(bh);
91 + ret = submit_bh(WRITE, bh);
93 - put_bh(bh); /* One for getblk() */
94 - journal_put_journal_head(descriptor);
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 - return (ret == -EIO);
128 + while (commit_transaction->t_locked_list) {
129 + struct buffer_head *bh;
131 + jh = commit_transaction->t_locked_list->b_tprev;
134 + if (buffer_locked(bh)) {
135 + spin_unlock(&journal->j_list_lock);
136 + wait_on_buffer(bh);
137 + if (unlikely(!buffer_uptodate(bh)))
139 + spin_lock(&journal->j_list_lock);
141 + if (!inverted_lock(journal, bh)) {
143 + spin_lock(&journal->j_list_lock);
146 + if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
147 + __journal_unfile_buffer(jh);
148 + jbd_unlock_bh_state(bh);
149 + journal_remove_journal_head(bh);
152 + jbd_unlock_bh_state(bh);
155 + cond_resched_lock(&journal->j_list_lock);
160 void journal_do_submit_data(struct buffer_head **wbuf, int bufs)
161 @@ -273,6 +350,20 @@ write_out_data:
162 journal_do_submit_data(wbuf, bufs);
165 +static inline __u32 jbd_checksum_data(__u32 crc32_sum, struct buffer_head *bh)
167 + struct page *page = bh->b_page;
171 + addr = kmap_atomic(page, KM_USER0);
172 + checksum = crc32_be(crc32_sum,
173 + (void *)(addr + offset_in_page(bh->b_data)),
175 + kunmap_atomic(addr, KM_USER0);
180 * journal_commit_transaction
182 @@ -296,6 +387,8 @@ void journal_commit_transaction(journal_
186 + struct buffer_head *cbh = NULL; /* For transactional checksums */
187 + __u32 crc32_sum = ~0;
190 * First job: lock down the current transaction and wait for
191 @@ -439,39 +532,14 @@ void journal_commit_transaction(journal_
192 journal_submit_data_buffers(journal, commit_transaction);
195 - * Wait for all previously submitted IO to complete.
196 + * Wait for all previously submitted IO to complete if commit
197 + * record is to be written synchronously.
199 spin_lock(&journal->j_list_lock);
200 - while (commit_transaction->t_locked_list) {
201 - struct buffer_head *bh;
203 - jh = commit_transaction->t_locked_list->b_tprev;
206 - if (buffer_locked(bh)) {
207 - spin_unlock(&journal->j_list_lock);
208 - wait_on_buffer(bh);
209 - spin_lock(&journal->j_list_lock);
211 - if (unlikely(!buffer_uptodate(bh)))
214 - if (!inverted_lock(journal, bh)) {
216 - spin_lock(&journal->j_list_lock);
219 - if (buffer_jbd(bh) && jh->b_jlist == BJ_Locked) {
220 - __journal_unfile_buffer(jh);
221 - jbd_unlock_bh_state(bh);
222 - journal_remove_journal_head(bh);
225 - jbd_unlock_bh_state(bh);
228 - cond_resched_lock(&journal->j_list_lock);
230 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
231 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT))
232 + err = journal_wait_on_locked_list(journal,
233 + commit_transaction);
234 spin_unlock(&journal->j_list_lock);
237 @@ -643,6 +712,16 @@ void journal_commit_transaction(journal_
239 for (i = 0; i < bufs; i++) {
240 struct buffer_head *bh = wbuf[i];
242 + * Compute checksum.
244 + if (JFS_HAS_COMPAT_FEATURE(journal,
245 + JFS_FEATURE_COMPAT_CHECKSUM)) {
247 + jbd_checksum_data(crc32_sum,
252 clear_buffer_dirty(bh);
253 set_buffer_uptodate(bh);
254 @@ -659,6 +738,23 @@ start_journal_io:
258 + /* Done it all: now write the commit record asynchronously. */
260 + if (JFS_HAS_INCOMPAT_FEATURE(journal,
261 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
262 + err = journal_submit_commit_record(journal, commit_transaction,
265 + __journal_abort_hard(journal);
267 + spin_lock(&journal->j_list_lock);
268 + err = journal_wait_on_locked_list(journal,
269 + commit_transaction);
270 + spin_unlock(&journal->j_list_lock);
272 + __journal_abort_hard(journal);
275 /* Lo and behold: we have just managed to send a transaction to
276 the log. Before we can commit it, wait for the IO so far to
277 complete. Control buffers being written are on the
278 @@ -757,9 +853,15 @@ wait_for_iobuf:
281 jbd_debug(3, "JBD: commit phase 6\n");
283 - if (journal_write_commit_record(journal, commit_transaction))
286 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
287 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)) {
288 + err = journal_submit_commit_record(journal, commit_transaction,
291 + __journal_abort_hard(journal);
293 + err = journal_wait_on_commit_record(cbh);
296 __journal_abort_hard(journal);
297 Index: linux-2.6.18.8/include/linux/jbd.h
298 ===================================================================
299 --- linux-2.6.18.8.orig/include/linux/jbd.h
300 +++ linux-2.6.18.8/include/linux/jbd.h
301 @@ -148,6 +148,29 @@ typedef struct journal_header_s
308 +#define JFS_CRC32_CHKSUM 1
309 +#define JFS_MD5_CHKSUM 2
310 +#define JFS_SHA1_CHKSUM 3
312 +#define JFS_CRC32_CHKSUM_SIZE 4
314 +#define JFS_CHECKSUM_BYTES (32 / sizeof(u32))
316 + * Commit block header for storing transactional checksums:
318 +struct commit_header
321 + __be32 h_blocktype;
323 + unsigned char h_chksum_type;
324 + unsigned char h_chksum_size;
325 + unsigned char h_padding[2];
326 + __be32 h_chksum[JFS_CHECKSUM_BYTES];
330 * The block tag: used to describe a single buffer in the journal
331 @@ -234,12 +257,16 @@ typedef struct journal_superblock_s
332 ((j)->j_format_version >= 2 && \
333 ((j)->j_superblock->s_feature_incompat & cpu_to_be32((mask))))
335 -#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
336 +#define JFS_FEATURE_COMPAT_CHECKSUM 0x00000001
338 +#define JFS_FEATURE_INCOMPAT_REVOKE 0x00000001
339 +#define JFS_FEATURE_INCOMPAT_ASYNC_COMMIT 0x00000004
341 /* Features known to this kernel version: */
342 -#define JFS_KNOWN_COMPAT_FEATURES 0
343 +#define JFS_KNOWN_COMPAT_FEATURES JFS_FEATURE_COMPAT_CHECKSUM
344 #define JFS_KNOWN_ROCOMPAT_FEATURES 0
345 -#define JFS_KNOWN_INCOMPAT_FEATURES JFS_FEATURE_INCOMPAT_REVOKE
346 +#define JFS_KNOWN_INCOMPAT_FEATURES (JFS_FEATURE_INCOMPAT_REVOKE | \
347 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)
351 @@ -1053,6 +1080,8 @@ extern int journal_check_available_fe
352 (journal_t *, unsigned long, unsigned long, unsigned long);
353 extern int journal_set_features
354 (journal_t *, unsigned long, unsigned long, unsigned long);
355 +extern int journal_clear_features
356 + (journal_t *, unsigned long, unsigned long, unsigned long);
357 extern int journal_create (journal_t *);
358 extern int journal_load (journal_t *journal);
359 extern void journal_destroy (journal_t *);
360 Index: linux-2.6.18.8/fs/jbd/recovery.c
361 ===================================================================
362 --- linux-2.6.18.8.orig/fs/jbd/recovery.c
363 +++ linux-2.6.18.8/fs/jbd/recovery.c
365 #include <linux/jbd.h>
366 #include <linux/errno.h>
367 #include <linux/slab.h>
368 +#include <linux/crc32.h>
372 @@ -307,6 +308,38 @@ int journal_skip_recovery(journal_t *jou
377 + * calc_chksums calculates the checksums for the blocks described in the
378 + * descriptor block.
380 +static int calc_chksums(journal_t *journal, struct buffer_head *bh,
381 + unsigned long *next_log_block, __u32 *crc32_sum)
383 + int i, num_blks, err;
384 + unsigned long io_block;
385 + struct buffer_head *obh;
387 + num_blks = count_tags(bh, journal->j_blocksize);
388 + /* Calculate checksum of the descriptor block. */
389 + *crc32_sum = crc32_be(*crc32_sum, (void *)bh->b_data, bh->b_size);
391 + for (i = 0; i < num_blks; i++) {
392 + io_block = (*next_log_block)++;
393 + wrap(journal, *next_log_block);
394 + err = jread(&obh, journal, io_block);
396 + printk(KERN_ERR "JBD: IO error %d recovering block "
397 + "%lu in log\n", err, io_block);
400 + *crc32_sum = crc32_be(*crc32_sum, (void *)obh->b_data,
408 static int do_one_pass(journal_t *journal,
409 struct recovery_info *info, enum passtype pass)
411 @@ -318,6 +351,7 @@ static int do_one_pass(journal_t *journa
412 struct buffer_head * bh;
413 unsigned int sequence;
415 + __u32 crc32_sum = ~0; /* Transactional Checksums */
417 /* Precompute the maximum metadata descriptors in a descriptor block */
418 int MAX_BLOCKS_PER_DESC;
419 @@ -409,9 +443,24 @@ static int do_one_pass(journal_t *journa
421 case JFS_DESCRIPTOR_BLOCK:
422 /* If it is a valid descriptor block, replay it
423 - * in pass REPLAY; otherwise, just skip over the
424 - * blocks it describes. */
425 + * in pass REPLAY; if journal_checksums enabled, then
426 + * calculate checksums in PASS_SCAN, otherwise,
427 + * just skip over the blocks it describes. */
428 if (pass != PASS_REPLAY) {
429 + if (pass == PASS_SCAN &&
430 + JFS_HAS_COMPAT_FEATURE(journal,
431 + JFS_FEATURE_COMPAT_CHECKSUM) &&
432 + !info->end_transaction) {
433 + if (calc_chksums(journal, bh,
444 count_tags(bh, journal->j_blocksize);
445 wrap(journal, next_log_block);
446 @@ -506,9 +555,97 @@ static int do_one_pass(journal_t *journa
449 case JFS_COMMIT_BLOCK:
450 - /* Found an expected commit block: not much to
451 - * do other than move on to the next sequence
452 + /* How to differentiate between interrupted commit
453 + * and journal corruption ?
455 + * {nth transaction}
456 + * Checksum Verification Failed
458 + * ____________________
460 + * async_commit sync_commit
462 + * | GO TO NEXT "Journal Corruption"
465 + * {(n+1)th transanction}
467 + * _______|______________
469 + * Commit block found Commit block not found
471 + * "Journal Corruption" |
472 + * _____________|__________
474 + * nth trans corrupt OR nth trans
475 + * and (n+1)th interrupted interrupted
476 + * before commit block
477 + * could reach the disk.
478 + * (Cannot find the difference in above
479 + * mentioned conditions. Hence assume
480 + * "Interrupted Commit".)
483 + /* Found an expected commit block: if checksums
484 + * are present verify them in PASS_SCAN; else not
485 + * much to do other than move on to the next sequence
487 + if (pass == PASS_SCAN &&
488 + JFS_HAS_COMPAT_FEATURE(journal,
489 + JFS_FEATURE_COMPAT_CHECKSUM)) {
490 + int chksum_err, chksum_seen;
491 + struct commit_header *cbh =
492 + (struct commit_header *)bh->b_data;
493 + unsigned found_chksum =
494 + be32_to_cpu(cbh->h_chksum[0]);
496 + chksum_err = chksum_seen = 0;
498 + if (info->end_transaction) {
499 + printk(KERN_ERR "JBD: Transaction %u "
500 + "found to be corrupt.\n",
501 + next_commit_ID - 1);
506 + if (crc32_sum == found_chksum &&
507 + cbh->h_chksum_type == JFS_CRC32_CHKSUM &&
508 + cbh->h_chksum_size ==
509 + JFS_CRC32_CHKSUM_SIZE) {
511 + } else if (!(cbh->h_chksum_type == 0 &&
512 + cbh->h_chksum_size == 0 &&
513 + found_chksum == 0 &&
516 + * If fs is mounted using an old kernel and then
517 + * kernel with journal_chksum is used then we
518 + * get a situation where the journal flag has
519 + * checksum flag set but checksums are not
520 + * present i.e chksum = 0, in the individual
522 + * Hence to avoid checksum failures, in this
523 + * situation, this extra check is added.
529 + info->end_transaction = next_commit_ID;
531 + if (!JFS_HAS_INCOMPAT_FEATURE(journal,
532 + JFS_FEATURE_INCOMPAT_ASYNC_COMMIT)){
534 + "JBD: Transaction %u "
535 + "found to be corrupt.\n",
546 @@ -544,9 +681,10 @@ static int do_one_pass(journal_t *journa
547 * transaction marks the end of the valid log.
550 - if (pass == PASS_SCAN)
551 - info->end_transaction = next_commit_ID;
553 + if (pass == PASS_SCAN) {
554 + if (!info->end_transaction)
555 + info->end_transaction = next_commit_ID;
557 /* It's really bad news if different passes end up at
558 * different places (but possible due to IO errors). */
559 if (info->end_transaction != next_commit_ID) {
560 Index: linux-2.6.18.8/fs/jbd/journal.c
561 ===================================================================
562 --- linux-2.6.18.8.orig/fs/jbd/journal.c
563 +++ linux-2.6.18.8/fs/jbd/journal.c
564 @@ -67,6 +67,7 @@ EXPORT_SYMBOL(journal_update_format);
565 EXPORT_SYMBOL(journal_check_used_features);
566 EXPORT_SYMBOL(journal_check_available_features);
567 EXPORT_SYMBOL(journal_set_features);
568 +EXPORT_SYMBOL(journal_clear_features);
569 EXPORT_SYMBOL(journal_create);
570 EXPORT_SYMBOL(journal_load);
571 EXPORT_SYMBOL(journal_destroy);
572 @@ -1573,6 +1574,33 @@ int journal_set_features (journal_t *jou
577 + * int journal_clear_features () - Clear a given journal feature in the superblock
578 + * @journal: Journal to act on.
579 + * @compat: bitmask of compatible features
580 + * @ro: bitmask of features that force read-only mount
581 + * @incompat: bitmask of incompatible features
583 + * Clear a given journal feature as present on the
584 + * superblock. Returns true if the requested features could be reset.
587 +int journal_clear_features (journal_t *journal, unsigned long compat,
588 + unsigned long ro, unsigned long incompat)
590 + journal_superblock_t *sb;
592 + jbd_debug(1, "Clear features 0x%lx/0x%lx/0x%lx\n",
593 + compat, ro, incompat);
595 + sb = journal->j_superblock;
597 + sb->s_feature_compat &= ~cpu_to_be32(compat);
598 + sb->s_feature_ro_compat &= ~cpu_to_be32(ro);
599 + sb->s_feature_incompat &= ~cpu_to_be32(incompat);
605 * int journal_update_format () - Update on-disk journal structure.
606 Index: linux-2.6.18.8/fs/Kconfig
607 ===================================================================
608 --- linux-2.6.18.8.orig/fs/Kconfig
609 +++ linux-2.6.18.8/fs/Kconfig
610 @@ -140,6 +140,7 @@ config EXT3_FS_SECURITY
616 This is a generic journaling layer for block devices. It is
617 currently used by the ext3 and OCFS2 file systems, but it could
618 Index: linux-2.6.18.8/Documentation/filesystems/ext3.txt
619 ===================================================================
620 --- linux-2.6.18.8.orig/Documentation/filesystems/ext3.txt
621 +++ linux-2.6.18.8/Documentation/filesystems/ext3.txt
622 @@ -14,6 +14,16 @@ Options
623 When mounting an ext3 filesystem, the following option are accepted:
626 +journal_checksum Enable checksumming of the journal transactions.
627 + This will allow the recovery code in e2fsck and the
628 + kernel to detect corruption in the kernel. It is a
629 + compatible change and will be ignored by older kernels.
631 +journal_async_commit Commit block can be written to disk without waiting
632 + for descriptor blocks. If enabled older kernels cannot
633 + mount the device. This will enable 'journal_checksum'
636 journal=update Update the ext3 file system's journal to the current