Install grub in template's root.img

[marmarek]

This is required to boot a template (or an AppVM based on it) as HVM - #2185 .
In our current root.img layout there is no partition table and not enough space before filesystem start, so it's tricky to get grub working there. Technically it's possible using not-recommended "blocklist" method, but better add a partition table and move filesystem to the first partition, leaving enough space before for grub.

Another issue is that grub don't like dm-linear (if not set through LVM) and fail to resolve /dev/mapper/dmroot.

A checklist:

• move filesystem to a partition (template builder) - do this optionally, to still support building old template formats
• actually install grub there
• look for root filesystem on the first partition (while constructing dmroot in initramfs)
• adjust fstab if needed
• adjust /etc/default/grub to disable disk UUID usage - it's unreliable when filesystem will be visible as both /dev/xvda1 (possibly read-only device) and /dev/mapper/dmroot (always read-write device - in AppVM with a CoW layer)

[marmarek]

As for grub not liking /dev/mapper/dmroot - since in TemplateVM (where grub will be installed) /dev/xvda (and /dev/xvda1) is read-write, maybe we should mount this one directly, instead of /dev/mapper/dmroot? The problem is it will not match root= parameter nor /etc/fstab (and we can't change those, as it will also affect AppVM - where dmroot should still be used). But maybe it isn't a problem?

[marmarek]

As for grub not liking /dev/mapper/dmroot - since in TemplateVM (where grub will be installed) /dev/xvda (and /dev/xvda1) is read-write, maybe we should mount this one directly, instead of /dev/mapper/dmroot? The problem is it will not match root= parameter nor /etc/fstab (and we can't change those, as it will also affect AppVM - where dmroot should still be used). But maybe it isn't a problem?

[marmarek]

Interesting thing - when /dev/mapper/dmroot is set using dm-linear on /dev/xvda1 (with offset 0), grub2-install fails (at least on Fedora 23). But if it's dm-linear on /dev/xvda (with offset pointing at the first partition - 2048), then it's ok. Very strange. And also dmroot setup looks much more non-obvious than before :/

[marmarek]

Interesting thing - when /dev/mapper/dmroot is set using dm-linear on /dev/xvda1 (with offset 0), grub2-install fails (at least on Fedora 23). But if it's dm-linear on /dev/xvda (with offset pointing at the first partition - 2048), then it's ok. Very strange. And also dmroot setup looks much more non-obvious than before :/

[marmarek]

In preparation for PVHv2 it may be good idea to create also GPT and ESP (/boot/efi) - http://markmail.org/message/5ycdxlta2j77ttz2

[marmarek]

In preparation for PVHv2 it may be good idea to create also GPT and ESP (/boot/efi) - http://markmail.org/message/5ycdxlta2j77ttz2

[marmarek]

And maybe instead of installing grub (i386) there, use UEFI to load VM-provided kernel, right now (i.e. even before real PVHv2)? This would greatly simplify VM setup, and maybe also slightly improve startup time. At the cost of slightly more complex stubdomain - need to load ovmf instead of seabios.
This approach would mean the same template layout for both pre-PVHv2 and PVHv2 templates.

[marmarek]

And maybe instead of installing grub (i386) there, use UEFI to load VM-provided kernel, right now (i.e. even before real PVHv2)? This would greatly simplify VM setup, and maybe also slightly improve startup time. At the cost of slightly more complex stubdomain - need to load ovmf instead of seabios.
This approach would mean the same template layout for both pre-PVHv2 and PVHv2 templates.