Time for smarter hard disks?

[work in progress]

Should hard disks get smarter and start handling some filesystem tasks by themselves? This idea has been floating around in my mind for a while. Disks know their physical layout and could probably allocate disk space more efficiently than any filesystem, not to mention that doing so would offload a significant part of what the filesystem code does from the main CPU.

The basic idea goes like this:

• Disks would think in terms of “objects” that directly represent files⁽¹⁾. They wouldn't handle human-friendly file names and file hierarchies: that task would be left to an operating system⁽²⁾.
• Objects would be identified by a 32- or 64-bit number⁽³⁾; object 0 would be a virtual object that returns a list of allocated identifiers.
• Objects would support basic file operations: create, open, seek, read, write, truncate, erase^(4,5). Renaming and hard links would be the operating system's responsibility, of course, as we've decided the disk has no notion of file names.
• Objects should support sparse-file semantics.

Additional possibilities and food for thought include:

• a “grow-in-the-middle” file operation: this is hard to do on current disks because they are usually constrained to block-size boundaries, but might be doable on a smart disk;
• storing tagged per-object metadata directly on disk (I'm thinking Rock Ridge-like stuff here) to store file permissions, ownership information and whatever else you may need (not the file name, see above).

1. Directories are just special files.
2. Hopefully, a standard “file hierarchy system” would be developed, to allow interoperability between all operating systems.
3. Unix-like operating systems could probably use the identifier as the inode number.
4. Also, a “secure erase” would likely be useful.
5. Unix-like erase semantics? I.e. can you erase open files?

Copyright © Flavio Stanchina
July 22, 2009 at 00:48:27