lupin merge progress

[Colin Watson]

On Mon, Jul 30, 2007 at 10:47:47PM +0100, ago wrote:
> Looks much neater now than my implementation.

Glad to hear it.

> 1) The partman-auto-loop preseeding requires a disk+partition number.
> But the preseed file is generated from within windows (at least when
> setup.exe is used), and I do not know any good way to map a windows
> drive to a linux device from within windows (some windows expert out
> there might correct me on this). A workaround might be to also allow
> to preseed some other data to be used to find out the host partition
> and the loop file via heuristics.

OK; I noticed that some bits of lupin do this by looking for a named
file on the filesystem, so maybe we can do that? A bit fiddly, but it's
possible.

> 2) The hosting filesystem will be mostly mounted ro in normal
> circumstances, which means that it should be remounted rw, e.g. by
> using a /host entry in fstab and by modifying
> /etc/init.d/checkroot.sh.

Right.

> I am not sure whether you can change the fstype when remounting, in
> which case ntfs-3g has to be installed but it does not need to be
> within the initrd (which can use ntfs).

Changing the driver would involve unmounting and mounting again, rather
than just mount -o remount. I think it's much more straightforward just
to put ntfs-3g in the initramfs and use it right from the start, really.

> 3) Once grub is up, you should be aware of an issue with paths inside
> menu.lst, since menu.lst is used by grub4dos (where the kernel will be
> in c:/ubuntu/boot/vmlinuz) but menu.lst is generated within ubuntu
> where the path to the same file looks different (/boot/vmlinuz). At
> the moment this is taken care by
> lupin/devel/src/lupin-target/etc/init.d/cpkernel but there is probably
> a more elegant way to cope with it.

OK. I think it may well make sense for lupin to continue to exist to
take care of differences that are specific to running Linux inside a
Windows filesystem (as opposed to just loop-mounting in general).

> > Comments welcome! Is it possible to update lupin to make use of this, as
> > we discussed a while back?
> 
> I can do that but if lupin.postinst is taken care of and initrd can
> take a loop argument, there will not be much left of lupin...

I have nothing against lupin, but that was the general idea. :-) To the
extent that lupin consisted of patches and workarounds for parts of the
installer that could be applied in a generally safe way, I wanted to
merge those, and I fully expected that lupin would get a lot smaller as
a result. If there's not a lot left, that's good, because it will make
it easier to see the parts that are left.

(I'd like to see the alternate-iso parts of lupin made into a proper
udeb, by the way. That's necessary if those parts are going to stick
around for things like grub menu.lst munging.)

> What is left, on top of the above, is:
> 
> A) fix sendsigs to avoid killing userspace filesystems (I have noticed
> you have already submitted a patch)

Right. If I can figure out a way to test that adequately, I may well
upload it in advance of Phillip's NTFS work.

The only tricky bit is extracting the pid of the ntfs-3g process (which
should be done in the initramfs and then propagated to
/var/run/sendsigs.omit for use at shutdown). I was hoping not to have to
parse ps output for this, but that might be the most straightforward
approach.

> B) disable hibernation/suspend on loopmounted installations (mjg59
> might be able to fix suspend issues, but hibernation should not be
> feasible), see lupin/devel/src/lupin-target/DEBIAN/postinst

(Source packages don't customarily have upper-case DEBIAN directories,
by the way; those are normally constructed on the fly by programs called
from debian/rules.)

The conventional places to put that kind of thing would probably be in a
/usr/lib/finish-install.d script in lupin-udeb and a cloned-and-hacked
version of that in ubiquity/scripts/install.py.

> C) modify sysctl to make the filesystem more robust in case of hard
> reboots (those can damage the journal of the nested filesystem). The
> settings used at the moment (as suggested by szaka, ntfs-3g author)
> are:
> 
> vm.dirty_background_ratio=0
> vm.dirty_ratio=0
> vm.dirty_expire_centisecs=1
> vm.dirty_writeback_centisecs=1

lupin shouldn't modify /etc/sysctl.conf for this; that's a conffile and
scripts may not modify it because that may result in conffile prompts on
later upgrades. It would probably be better for there to be a
properly-packaged lupin.deb shipping an init script that does that.

> D) Patch os-prober and m-a. See
> lupin/devel/src/alternate-iso/overlay/var/lib/dpkg/info/load-cdrom.postinst

I've applied the os-prober change upstream. Evan, could you look at the
migration-assistant changes? It looks like some of them may be
generally-applicable and some may need further thought as to how to
merge them properly, but I've only glanced over the sed expressions and
not checked the code it's patching.

Cheers,

-- 
Colin Watson                                       [cjwatson at ubuntu.com]