File attributes

[martin at minster.york.ac.uk]

Andrew Hume said:
>	... Why not add one more system call (if you really hate this,
> then standardise a way in via fcntl), say getattr, that is analogous to
> stat and returns the value of an ascii attribute name? Then the files ...

I implemented a filing system (for storing sampled, CD-quality sounds)
which implements something like this - we call them `properties' not
attributes - a null-terminated textual key, is mapped to an arbitrary
binary value (with a stored length).  In practice many values turn out
to be textual, but the flexibility of binary values has been useful.

A single `getattr'-like call is not enough - one also needs to be able
to remove properties (we allow `binary properties' that have no value,
and whose significance is their presence or absence in the file, so a
zero value length cannot be used to remove a property. I suppose a
negative length could be used, but this sounds like a hack to me :-).
More interestingly one needs to be able to find out the keys of all the
properties defined for a file, so the copy program can perserve them,
and other more interesting manipulations can be carried out. In my
system this also returns the size of the key's value, so that the
application knows how much space to allocate if it has to read the
property.

Having done all this, properties/attributes are something of a mixed
blessing. Managing the namespace, and standardizing the properties that
are stored with files is a nightmare! Centralized administration is out
of the question (no resources to support this, it is too awkward for
developers and totally unenforceable) while locally defined property
names quickly lead to chaos! At other times properties can be quite
useful :-)!

In a new version of the system that I am currently designing/thinking
about I would prefer to use a perfectly good data agregation mechanism
that is already in Unix, MSDOS, TOS, etc. I am of course refering to
directories! Changes to allow directories to be moved and copied more
freely seem to be enough to make this a better solution than attaching
properties directly to the files. This would also allow `properties'
that were hard or symbolic links, and other obvious generalizations.
It also makes it much easier to avoid horrible restrictions/implementation
problems on the size (and the changes in size) of properties.

Martin
usenet: ...!mcsun!ukc!minster!martin
JANET:  martin at uk.ac.york.minster
INTERNET: martin%minster.york.ac.uk at nsfnet-relay.ac.uk
surface:
	Dr. Martin C. Atkins
	Dept. of Computer Science
	University of York
	Heslington
	York YO1 5DD
	ENGLAND