make CD images valid archives for netboot

[Yann Vernier]

I attempted a network install on my laptop as it has no CD drive. I grabbed the
kernel and initrd from install/netboot on the warty test iso. The installer got
stuck at trying to partition, and I found that the system contained no hint of
IDE existing on my machine - not even a mention of the "Uniform Multi-Platform
E-IDE driver". Output from dmesg and lspci on an installed system are available
at the URL.

http://donkey.dyndns.org/~yann/palmax/: http://donkey.dyndns.org/~yann/palmax/

[Yann Vernier (yann)]

Created an attachment (id=78)
Output from lspci -v

[Yann Vernier (yann)]

Created an attachment (id=79)
Output from dmesg for working kernel

[Matt Zimmerman (mdz)]

Sounds like the problem is likely with hw-detect; it's responsible for loading
kernel drivers in the installer. discover1 seems to have an entry in its database

[Yann Vernier (yann)]

(In reply to comment #3)
> Sounds like the problem is likely with hw-detect; it's responsible for loading
> kernel drivers in the installer. discover1 seems to have an entry in its database

I poked around in the installer using the shell on vt2, and there were no IDE
modules around except for ide-cs.

[Yann Vernier (yann)]

Okay, problem identified. The IDE module udeb packages were missing from the
ISO, and I was trying to install from a local machine having that loop mounted.
The udebs are present in the online archive. Please include all udebs on the
next CD.

[Colin Watson (cjwatson)]

No, it doesn't work that way. On the CD images, IDE modules are unpacked into
the initrd, since you may well need them in order to be able to read udebs from
the CD. Since they're already on the initrd, there's no need to include them as
udebs as well; it would just be a waste of space.

Conversely, the netboot initrds don't need IDE modules in order to read udebs,
since they retrieve them from the network.

If I understand you correctly, you tried to use a netboot initrd but a
loop-mounted CD as a hacked-up substitute for an archive; I'm afraid that isn't
a supported configuration (neither in Debian nor in Ubuntu), and we'd rather
reserve the required space for other things. I suggest using a real Ubuntu
archive instead. If you're short on bandwidth to a mirror, you can certainly use
the contents of the CD to seed the mirror, which would speed things up a lot.

We should probably document this, though.

[Yann Vernier (yann)]

We're talking about 170KB here. And you might say I'm "short on bandwidth" - one
of the target locations is basically a connectionless bunker. Personally I think
you could invest that space into making the CD+netboot initrd image included on
it work together. I understand that it's only a minor and hackish request, but
it doesn't need a large amount of resources.

[Matt Zimmerman (mdz)]

How much space would it cost us in order to make this work? We aren't short on
space at the moment, and it seems awfully convenient to be able to do this

[Colin Watson (cjwatson)]

I finally got round to doing a size analysis of this. Removing exclude-udebs*
completely, and thus causing everything in the installer seed to end up on the
CD as well as in any initrds they might be in, costs 3.2MB on i386 and 10.8MB on
powerpc (powerpc has three times the number of kernel modules and I think had
slightly more exclusions). This isn't as much as I'd expected, and I think we
can live with it. Apart from anything else, our exclude-udebs* files hadn't been
kept up to date and needed to be reworked anyway.

If I exclude kernel-image-*, which is easy, doesn't affect this bug, and
requires no maintenance, the cost goes down to 1.9MB on i386 and 5.5MB on powerpc.

Accordingly, I've arranged for all udebs in the main archive except
kernel-image-* to end up on the CD, starting from tomorrow's daily build; that
should fix this bug. Eventually we may be able to come up with something clever
to reduce size safely while taking this bug into account, but it doesn't seem
urgent.