summaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--fs/ubifs/tnc_commit.c26
1 files changed, 26 insertions, 0 deletions
diff --git a/fs/ubifs/tnc_commit.c b/fs/ubifs/tnc_commit.c
index f50c3e5..d6fab1a 100644
--- a/fs/ubifs/tnc_commit.c
+++ b/fs/ubifs/tnc_commit.c
@@ -89,6 +89,10 @@ static int make_idx_node(struct ubifs_info *c, struct ubifs_idx_node *idx,
ubifs_assert(ubifs_zn_dirty(znode));
ubifs_assert(ubifs_zn_cow(znode));
+ /*
+ * Note, unlike 'write_index()' we do not add memory barriers here
+ * because this function is called with @c->tnc_mutex locked.
+ */
__clear_bit(DIRTY_ZNODE, &znode->flags);
__clear_bit(COW_ZNODE, &znode->flags);
@@ -903,6 +907,28 @@ static int write_index(struct ubifs_info *c)
clear_bit(COW_ZNODE, &znode->flags);
smp_mb__after_clear_bit();
+ /*
+ * We have marked the znode as clean but have not updated the
+ * @c->clean_zn_cnt counter. If this znode becomes dirty again
+ * before 'free_obsolete_znodes()' is called, then
+ * @c->clean_zn_cnt will be decremented before it gets
+ * incremented (resulting in 2 decrements for the same znode).
+ * This means that @c->clean_zn_cnt may become negative for a
+ * while.
+ *
+ * Q: why we cannot increment @c->clean_zn_cnt?
+ * A: because we do not have the @c->tnc_mutex locked, and the
+ * following code would be racy and buggy:
+ *
+ * if (!ubifs_zn_obsolete(znode)) {
+ * atomic_long_inc(&c->clean_zn_cnt);
+ * atomic_long_inc(&ubifs_clean_zn_cnt);
+ * }
+ *
+ * Thus, we just delay the @c->clean_zn_cnt update until we
+ * have the mutex locked.
+ */
+
/* Do not access znode from this point on */
/* Update buffer positions */
OpenPOWER on IntegriCloud