UBIFS: add few commentaries about TNC

Add a couple of comments - while looking into TNC I could not easily figure out
few facts, so it is a good idea to document them in the code.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

diff --git a/fs/ubifs/tnc_commit.c b/fs/ubifs/tnc_commit.c
index f50c3e5da263830ee944d161c5b1dd6d355679d8..d6fab1a9986c38642b6815bb4c29bc5baff10c43 100644 (file)
--- 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 */