In this paper, we extend understanding of semantically-smart disks by presenting techniques to handle more general file system behaviors. Previous work required the file system to be mounted synchronously when implementing complex functionality within the disk; we relax that requirement. We now describe our assumptions for general file system behavior; we believe that many, if not all, modern file systems adhere to these behavioral guidelines.
First, blocks in a file system can be dynamically typed, i.e., the file system can locate different types of blocks at the same physical location on disk over the lifetime of the file system. For example, in a UNIX file system, a block in the data region can be a user-data block, an indirect-pointer block or a directory-data block. Second, a file system can delay updates to disk; delayed writes at the file system facilitate batching of small writes in memory and suppressing of writes to files that are subsequently deleted. Third, as a consequence of delayed writes, the order in which the file system actually writes data to disk can be arbitrary. Although certain file systems order writes carefully [14], to remain general, we do not make any such assumptions on the ordering. Note that our assumptions are made for practical reasons: the Linux ext2 file system exhibits all the aforementioned behaviors.