[Devel] [RH7 PATCH 2/2] jbd2: raid amnesia protection for the journal

[Dmitry Monakhov]

https://jira.sw.ru/browse/PSBM-15484

Some blockdevices can return different data on read requests from same block
after power failure (for example mirrored raid is out of sync, and resync is
in progress) In that case following sutuation is possible:

Power failure happen after transaction commit log was issued for
transaction 'D', next boot first dist will have commit block, but
second one will not.
mirror1: journal={Ac-Bc-Cc-Dc }
mirror2: journal={Ac-Bc-Cc-D  }
Now let's let assumes that we read from mirror1 and found that 'D' has
valid commit block, so journal_replay will replay that transaction, but
second power failure may happen before journal_reset() so next
journal_replay() may read from mirror2 and found that 'C' is last valid
transaction. This result in corruption because we already replayed
trandaction 'D'.
In order to avoid such ambiguity we should pefrorm 'stabilize write'.
1) Read and rewrite latest commit id block
2) Invalidate next block in
order to guarantee that journal head becomes stable.

Signed-off-by: Dmitry Monakhov <dmonakhov at openvz.org>
---
 fs/jbd2/recovery.c |   77 +++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 files changed, 76 insertions(+), 1 deletions(-)

diff --git a/fs/jbd2/recovery.c b/fs/jbd2/recovery.c
index 626846b..253485c 100644
--- a/fs/jbd2/recovery.c
+++ b/fs/jbd2/recovery.c
@@ -36,6 +36,9 @@ struct recovery_info
 	int		nr_replays;
 	int		nr_revokes;
 	int		nr_revoke_hits;
+
+	unsigned int		last_log_block;
+	struct buffer_head	*last_commit_bh;
 };
 
 enum passtype {PASS_SCAN, PASS_REVOKE, PASS_REPLAY};
@@ -233,6 +236,71 @@ do {									\
 		var -= ((journal)->j_last - (journal)->j_first);	\
 } while (0)
 
+/*
+ * The 'Raid amnesia' effect protection: https://jira.sw.ru/browse/PSBM-15484
+ *
+ * Some blockdevices can return different data on read requests from same block
+ * after power failure (for example mirrored raid is out of sync, and resync is
+ * in progress) In that case following sutuation is possible:
+ *
+ * Power failure happen after transaction commit log was issued for
+ * transaction 'D', next boot first dist will have commit block, but
+ * second one will not.
+ * mirror1: journal={Ac-Bc-Cc-Dc }
+ * mirror2: journal={Ac-Bc-Cc-D  }
+ * Now let's let assumes that we read from mirror1 and found that 'D' has
+ * valid commit block, so journal_replay will replay that transaction, but
+ * second power failure may happen before journal_reset() so next
+ * journal_replay() may read from mirror2 and found that 'C' is last valid
+ * transaction. This result in corruption because we already replayed
+ * trandaction 'D'.
+ * In order to avoid such ambiguity we should pefrorm 'stabilize write'.
+ * 1) Read and rewrite latest commit id block
+ * 2) Invalidate next block in
+ * order to guarantee that journal head becomes stable.
+ * Yes i know that 'stabilize write' approach is ugly but this is the only
+ * way to run filesystem on blkdevices with 'raid amnesia' effect
+ */
+static int stabilize_journal_head(journal_t *journal, struct recovery_info *info)
+{
+	struct buffer_head *bh[2] = {NULL, NULL};
+	int err, err2, i;
+
+	if (!info->last_commit_bh)
+		return 0;
+
+	bh[0] = info->last_commit_bh;
+	info->last_commit_bh = NULL;
+
+	err = jread(&bh[1], journal, info->last_log_block);
+	if (err)
+		goto out;
+
+	for (i = 0; i < 2; i++) {
+		lock_buffer(bh[i]);
+		/* Explicitly invalidate block beyond last commit block */
+		if (i == 1)
+			memset(bh[i]->b_data, 0, journal->j_blocksize);
+
+		BUFFER_TRACE(bh[i], "marking dirty");
+		set_buffer_uptodate(bh[i]);
+		mark_buffer_dirty(bh[i]);
+		BUFFER_TRACE(bh[i], "marking uptodate");
+		unlock_buffer(bh[i]);
+	}
+	err = sync_blockdev(journal->j_dev);
+	/* Make sure data is on permanent storage */
+	if (journal->j_flags & JBD2_BARRIER) {
+		err2 = blkdev_issue_flush(journal->j_dev, GFP_KERNEL, NULL);
+		if (!err)
+			err = err2;
+	}
+out:
+	brelse(bh[0]);
+	brelse(bh[1]);
+	return err;
+}
+
 /**
  * jbd2_journal_recover - recovers a on-disk journal
  * @journal: the journal to recover
@@ -270,6 +338,8 @@ int jbd2_journal_recover(journal_t *journal)
 
 	err = do_one_pass(journal, &info, PASS_SCAN);
 	if (!err)
+		err = stabilize_journal_head(journal, &info);
+	if (!err)
 		err = do_one_pass(journal, &info, PASS_REVOKE);
 	if (!err)
 		err = do_one_pass(journal, &info, PASS_REPLAY);
@@ -319,6 +389,7 @@ int jbd2_journal_skip_recovery(journal_t *journal)
 	memset (&info, 0, sizeof(info));
 
 	err = do_one_pass(journal, &info, PASS_SCAN);
+	brelse(info.last_commit_bh);
 
 	if (err) {
 		printk(KERN_ERR "JBD2: error %d scanning journal\n", err);
@@ -418,6 +489,7 @@ static int do_one_pass(journal_t *journal,
 {
 	unsigned int		first_commit_ID, next_commit_ID;
 	unsigned long		next_log_block;
+	unsigned long		last_commit_block;
 	int			err, success = 0;
 	journal_superblock_t *	sb;
 	journal_header_t *	tmp;
@@ -437,6 +509,7 @@ static int do_one_pass(journal_t *journal,
 	sb = journal->j_superblock;
 	next_commit_ID = be32_to_cpu(sb->s_sequence);
 	next_log_block = be32_to_cpu(sb->s_start);
+	last_commit_block = 0;
 
 	first_commit_ID = next_commit_ID;
 	if (pass == PASS_SCAN)
@@ -749,7 +822,9 @@ static int do_one_pass(journal_t *journal,
 					break;
 				}
 			}
-			brelse(bh);
+			brelse(info->last_commit_bh);
+			info->last_commit_bh = bh;
+			info->last_log_block = next_log_block;
 			next_commit_ID++;
 			continue;
 
-- 
1.7.1