bcache: fix a bug with the journal seq blacklist machinery

2 files changed, 16 insertions, 1 deletions

diff --git a/drivers/md/bcache/journal.c b/drivers/md/bcache/journal.c
index 6c3cc983dd79..a1b08e1c3ef1 100644
--- a/drivers/md/bcache/journal.c
+++ b/drivers/md/bcache/journal.c
@@ -331,7 +331,11 @@ int bch_journal_seq_should_ignore(struct cache_set *c, u64 seq, struct btree *b)
 	if (seq <= j->seq)
 		goto out;
 
-	cache_set_inconsistent_on(seq > j->seq + 2, c,
+	/*
+	 * Decrease this back to j->seq + 2 when we next rev the on disk format:
+	 * increasing it temporarily to work around bug in old kernels
+	 */
+	cache_set_inconsistent_on(seq > j->seq + 4, c,
 			 "bset journal seq too far in the future: %llu > %llu",
 			 seq, j->seq);
 
diff --git a/drivers/md/bcache/super.c b/drivers/md/bcache/super.c
index 69b44f81495c..e790503b9a92 100644
--- a/drivers/md/bcache/super.c
+++ b/drivers/md/bcache/super.c
@@ -1344,6 +1344,17 @@ static const char *run_cache_set(struct cache_set *c)
 
 		bch_verbose(c, "journal replay done");
 
+		/*
+		 * Write a new journal entry _before_ we start journalling new
+		 * data - otherwise, we could end up with btree node bsets with
+		 * journal seqs arbitrarily far in the future vs. the most
+		 * recently written journal entry on disk, if we crash before
+		 * writing the next journal entry:
+		 */
+		err = "error writing journal entry";
+		if (bch_journal_meta(&c->journal))
+			goto err;
+
 		bch_verbose(c, "starting fs gc:");
 		err = "error in fs gc";
 		ret = bch_gc_inode_nlinks(c);