Our assumptions about general file system behavior imply that the storage system cannot accurately classify the type of each block. Block classification is straightforward when the type of the block depends upon its location on disk. For example, in the Berkeley Fast File System (FFS) [28], the regions of disk that store inodes are fixed at file system creation; thus, any traffic to those regions is known to contain inodes.
However, type information is sometimes spread across multiple blocks. For example, a block filled with indirect pointers can only be identified as such by observing the corresponding inode, specifically that the inode's indirect pointer field contains the address of the given indirect block. More formally, to identify an indirect block B, the semantic disk must look for the inode that has block B in its indirect pointer field. Thus, when the relevant inode block IB is written to disk, the disk infers that B is an indirect block, and when it later observes block B written, it uses this information to classify and treat the block as an indirect block. However, due to the delayed write and reordering behavior of the file system, it is possible that in the time between the disk writes of IB and B, block B was freed from the original inode and was reallocated to another inode with a different type, i.e., as a normal data block. The disk does not know this since the operations took place in memory and were not reflected to disk. Thus, the inference made by the semantic disk on the block type could be wrong due to the inherent staleness of the information tracked. Implementing a correct system despite potentially inaccurate inferences is one of the challenges we address in this paper.