Subject: Re: union filesystem problems
To: Chris G. Demetriou <email@example.com>
From: John F. Woods <firstname.lastname@example.org>
Date: 05/11/1999 13:46:43
> In my opinion, the practice of putting everthing under the sun into
> GENERIC just because it compiles and maybe mostly works is broken. As
> far as I'm concerned, GENERIC should be "Enough to get booted, with
> support for all of the devices likely to be in the machine, but not
> much more."
Two ideas occur to me about this:
First, it's certainly useful to have a well-tested "lamp-test" kernel
with everything under the sun thrown into it, to ensure that there are
no surprising unpleasant interactions (there will always be unpleasant
interactions, probably, but you don't want them to be surprises). You
could make that a developer-only test kernel, but then you are relying
on having enough developers with enough hardware variety to actually
exercise everything; until NetBSD can afford a test lab with a couple
of hundred machines of each architecture, you'll have to rely on the
end-user to test the wierd corners (unfortunately).
Second, you want the vast majority of users to be able to just load the
install media and set themselves up, regardless of what wierd application
environment they want to have. But it could be argued that a forms-based
system configurator and a streamlined build process (especially if you
stick to precompiled .o files from a CD) is a better approach.
> Realistically, given their current level of functionality, what
> fraction of users find _any_ real utility from NULLFS, KERNFS, FDESC,
> PORTAL, UMAPFS, PROCFS, or UNION? How many users actually care about
the UFS QUOTA code, or XNS, CCITT, or ISO networking?
I certainly pitch most of those options in my config files, though I keep
KERNFS and FDESC. KERNFS and PROCFS would be a lot more useful if, well,
they got used; instead, however, there is an armed camp of developers who
won't use them because they aren't of any use for debugging a paper-tape
core dump of a dead system, which means those features can't be depended
on even by people who are not writing necromantic divinatory tools.
> It's not an insignificant number, but i'd bet there's a high correlation
> between that set of users and the set that has a strong understanding of
> what to put in a kernel config.
Hmm. Anyone who thinks they need ISO networking obviously understands nothing
about networks, are you sure they'll understand kernels any better? <ducks>