[libvirt-users] uefi built from tiancore via edk2 can't persist boot changes

Thu Mar 24 20:57:48 UTC 2016

On 03/21/16 19:53, jsl6uy js16uy wrote:
> Apologies if this has been gone over, but I believe I have checked the
> intertubes more than a bit.....
> 
> I am using libvirt and have vms booting under an OVMF.fd to use an efi
> firmware. I can create vms, linux ubuntu, and they will boot up.
> However, everytime I reboot am I dropped into the default efi shell
> provide by the tianocore build.
> Then I must walk the FS to the booting efi app and run, in this case
> grubx64.efi, to actually finish booting the host.
> I had tried adding boot entries with efibootmgr within the OS and also
> bcfg with the efi shell. I get no errors when adding an entry, and the
> new entry shows up and I can manipulate the entry i.e. set it to next
> boot and the like. Soon as I reboot I get dumped back to efi shell. 
> 
> Am I missing something? Are the var changes not being stored? Is there
> someplace to look for an error perhaps?
> 
> I also can use the same efi bootloader under libvirt to boot a hybrid
> iso. location of the efi application on the hybrid iso is
> /EFI/boot/bootx64.efi. Drops me right into my grub menu, after falling
> thru from EFI FLOPPY1 and EFI FLOPPY2 to EFI DVD
> So since that was/is working,  tried adding that to my efi partition so
> it has /EFI/boot/bootx64.efi. The thinking being that since I can't add
> an entry, I will setup for the "known" efi boot path. Of course that
> didn't work either
> 
> I am on arch linux:
> 
> Linux X 4.4.5-1-ARCH #1 SMP PREEMPT Thu Mar 10 07:38:19 CET 2016 x86_64
> GNU/Linux
> local/libvirt 1.3.2-3
>     API for controlling virtualization engines
> (openvz,kvm,qemu,virtualbox,xen,etc)
> local/libvirt-python 1.3.1-1
>     libvirt python binding
> 
> UEFI:
> Shell> ver
> UEFI Interactive Shell v2.1
> EDK II
> UEFI v2.50 (EDK II, 0x00010000)
> 
> 
> for this host....my nvram setting look like
>   <os>
>     <type arch='x86_64' machine='pc-i440fx-2.5'>hvm</type>
>     <loader readonly='yes' type='pflash'>/home/xyz/OVMF.fd</loader>
>     <nvram
> template='/usr/share/OVMF/OVMF_VARS.fd'>/var/lib/libvirt/qemu/nvram/X_VARS.fd</nvram>
>     <boot dev='hd'/>
>   </os>

Your domain XML is wrong. You can use OVMF with libvirt in three setups.

(1) worst setup: map the unified OVMF.fd file (which includes variable
store and fimrware executable) as ROM. I don't want people to use this,
ever, so although libvirt does support it, I'm not going to share the
XML snippet that implements this.

(2) a bit less wrong setup (but still pretty wrong): map the unified
OVMF.fd file as writeable pflash. This is bad because (a) you can't
share the unified OVMF.fd file between guests, and (b) whenever you
rebuild the OVMF.fd file, the guest that is pointed to it will lose all
of its persistent variables.

(3) the right setup: you almost got it right, but the <loader> element
in this setup must not point to the unified OVMF.fd file. It must point
to the executable-only part, the OVMF_CODE.fd file. So here's what you need:

  <os>
    <loader
     readonly='yes'
     type='pflash'
    >/your/build/output/OVMF_CODE.fd</loader>
    <nvram
     template='/your/build/output/OVMF_VARS.fd'
    >/var/lib/libvirt/qemu/nvram/guest_name_VARS.fd</nvram>
  </os>

This allows the firmware binary to be upgraded centrally, while every VM
can keep its private varstore.

The virt-install(1) manual describes this, under the example

--boot
loader=/.../OVMF_CODE.fd,loader_ro=yes,loader_type=pflash,nvram_template=/.../OVMF_VARS.fd

The way you specified it implies that the varstore range will be mapped
as read-only pflash directly from the unified OVMF.fd file, which
happens to make the varstore non-writeable. In addition, the real
varstore file will be mapped below *that*, which is completely bogus.

(4) Instead of <boot> elements in the <os> element, <boot order='...'/>
elements are preferred under <disk> and friends.

(5) You can (and are encouraged to) look at the OVMF log. If you want to
capture it, you should implement the following changes:

- add the following attribute to the root element of the domain XML:

  <domain
   type='kvm'
   xmlns:qemu='http://libvirt.org/schemas/domain/qemu/1.0'  <-- this one
  >

- add the following XML fragment near the end of the domain XML, as a
direct child of the <domain> element:

    <qemu:arg value='-global'/>
    <qemu:arg value='isa-debugcon.iobase=0x402'/>
    <qemu:arg value='-debugcon'/>
    <qemu:arg value='file:/tmp/guest_name.log'/>

This will send the OVMF log to /tmp/guest_name.log.

(6) If you are already missing a boot option to boot your installed OS,
you might have to reestablish that boot option manually (after the above
changes) -- add the option, and place it first in the boot order. You
can use efibootmgr for this, or the BCFG command, or simply the firmware
TUI.

I could go into why, but this time I won't. If you are interested, you
can mostly find the reason in the OVMF whitepaper
<http://people.redhat.com/~lersek/ovmf-whitepaper-c770f8c.txt>, section
"Platform-specific boot policy", bullet (b), "Auto-generation of boot
options". The point is, boot options pointing to arbitrary boot loader
UEFI applications on non-removable media cannot be autogenerated, hence
OVMF cannot filter them in. If you install the guest OS with the correct
domain XML in the first place, then things will work.

Some UEFI OSes also install "system rescue" utilities to the target
non-removable media, under the default removable media pathname. The
idea is that if you lose your "main" boot option (pointing to an
arbitrary OS boot loader), then the firmware can still auto-generate a
boot option for this fallback utility, and the fallback utility can
restore the OS-specific boot loader's boot option. You can read more
about this at
<https://blog.uncooperative.org/blog/2014/02/06/the-efi-system-partition/>.

Summary:
- use a domain XML like described in (3) and (4)
- when in doubt, capture the OVMF debug log as described in (5)

Thanks
Laszlo