Too Busy For Words - the PaulWay Blog

I was thinking last night, as I vainly searched for sleep, about a long-standing idea of mine: the Blockless File System. If you imagine the entire disk as just a big string of bytes, then several problems go away. You don't need to have special ways of keeping small files or tail-ends of files in sub-parts of blocks (or, for that matter, waste the half-block (on average) at the end of files that aren't a neat multiple of the block length). The one I'm really excited about is the ability to insert and delete arbitrary lengths within a file. Or append to the start of a file. And what about file versioning?

Aaaanyway, for some reason I was thinking of what would go in the superblock. I have only done a tiny bit of study into what goes into the superblock in modern file systems, so I'm not speaking from the point of view of a learned expert. But I thought one idea would be for a header that could detect which endian-ness the file system was written in. A quick five seconds of thought produced the idea that the block 'HLhl' would not only be fairly easy to recognise in a binary, bit-oriented way, but would make it really easy for a human to check that it was big-endian ('HLhl'), little-endian ('lhLH'), middle-endian ('hlHL') or any of the twenty-one other combinations of 32-bit endianness the computing world has yet to explore.

It would also mean that the reader could simply translate whatever they read into their own endian-ness and then write that straight to disk (including setting the HLhl header in their native endian-ness). So writes are not penalised and reads are only penalised once. If the entire disk was written this way, it would mean that you could take the device to a machine with different endian-ness and it would behave almost exactly like normal - almost no loss of speed in writing or reading little-endian defined inode numbers or what-have-you. The entire disk could be a mix of endian-ness styles and it would still be perfectly readable.

Of course, it's not really going to matter if you don't take your disks to foreign architectures, or those architectures support some sort of SWAB instruction to swap endianness with little slow down, or if the users of disks that have been taken to foreign architectures don't mind a bit of slowdown reading the file system structure.

This is probably one to chalk up in the "Paul solves the problems of the early 1970s computer industry - in 2006" board.