[JFFS2] Fix free space leak with in-band cleanmarkers

We were accounting for the cleanmarker by calling jffs2_link_node_ref()
(without locking!), which adjusted both superblock and per-eraseblock
accounting, subtracting the size of the cleanmarker from {jeb,c}->free_size
and adding it to {jeb,c}->used_size.

But only _then_ were we adding the size of the newly-erased block back
to the superblock counts, and we were adding each of jeb->{free,used}_size
to the corresponding superblock counts. Thus, the size of the cleanmarker
was effectively subtracted from the superblock's free_size _twice_.

Fix this, by always adding a full eraseblock size to c->free_size when
we've erased a block. And call jffs2_link_node_ref() under the proper
lock, while we're at it.

Thanks to Alexander Yurchenko and/or Damir Shayhutdinov for (almost)
pinpointing the problem.
Signed-off-by: David Woodhouse <dwmw2@infradead.org>

fs/jffs2/erase.c

@@ -419,9 +419,6 @@ static void jffs2_mark_erased_block(struct jffs2_sb_info *c, struct jffs2_eraseb
 			if (jffs2_write_nand_cleanmarker(c, jeb))
 				goto filebad;
 		}
-
-		/* Everything else got zeroed before the erase */
-		jeb->free_size = c->sector_size;
 	} else {
 
 		struct kvec vecs[1];
@@ -449,18 +446,19 @@ static void jffs2_mark_erased_block(struct jffs2_sb_info *c, struct jffs2_eraseb
 
 			goto filebad;
 		}
-
+	}
 	/* Everything else got zeroed before the erase */
 	jeb->free_size = c->sector_size;
-		/* FIXME Special case for cleanmarker in empty block */
-		jffs2_link_node_ref(c, jeb, jeb->offset | REF_NORMAL, c->cleanmarker_size, NULL);
-	}
 
 	mutex_lock(&c->erase_free_sem);
 	spin_lock(&c->erase_completion_lock);
+
 	c->erasing_size -= c->sector_size;
-	c->free_size += jeb->free_size;
-	c->used_size += jeb->used_size;
+	c->free_size += c->sector_size;
+
+	/* Account for cleanmarker now, if it's in-band */
+	if (c->cleanmarker_size && !jffs2_cleanmarker_oob(c))
+		jffs2_link_node_ref(c, jeb, jeb->offset | REF_NORMAL, c->cleanmarker_size, NULL);
 
 	jffs2_dbg_acct_sanity_check_nolock(c,jeb);
 	jffs2_dbg_acct_paranoia_check_nolock(c, jeb);