Re: booting btrfs

[Chris Murphy]

On Jan 9, 2014, at 3:03 AM, Michael Chang <address@hidden> wrote:

> 2014/1/9 Chris Murphy <address@hidden>:
>> 
>> On Jan 7, 2014, at 10:55 AM, Chris Murphy <address@hidden> wrote:
>> 
>>> 
>>> On Jan 1, 2014, at 10:17 PM, Michael Chang <address@hidden> wrote:
>>>> 
>>>> We snapshot /boot for kernel and initrd, otherwise the rollback would
>>>> encounter problem of incompatible userland and kernel/kernel modules.
>>>> And we need the ability to rollback them in terms of usefulness.
>>> 
>>> Of course, understood.
>>> 
>>> core.img is only going to point to one /boot, which may not be the /boot 
>>> snapshot needed for the kernel and initrd. This will be really confusing 
>>> for mortal users. They'd be unlikely to figure it out and understand it, 
>>> without documentation.
>>> 
>>> If core.img points to the "current" /boot, which it should, that boot has 
>>> the accumulated knowledge of all snapshots, and any recently updated grub 
>>> modules. Choosing to boot a snapshot means using a different /boot for 
>>> kernel/initramfs than what grub is using for its root. I don't off hand see 
>>> a problem with this because it's literally just two files that grub loads 
>>> from a different boot subvolume, found with an absolute path to that 
>>> snapshot. And it also uses rootflags=subvol= to use the matching root 
>>> snapshot.
>>> 
>>> A separate issue that's not grub's problem is how to deal with the (now 
>>> wrong) fstab entries. systemd looks at fstab and generates mount jobs from 
>>> that; if taught to understand it's booting a snapshot it could second guess 
>>> parts of the fstab. Based on the name of the currently booting root 
>>> snapshot, which systemd definitely knows, it could mount that subvol= 
>>> instead of what's in fstab. It can use name substitution to do the same 
>>> thing with the other subvolume-snapshots that match the root one. Meaning 
>>> all of the snapshots for a system have the same base (re)naming convention.
>> 
>> Another hiccup. Maybe it's a silly use case. Consider /boot on Btrfs, 
>> multiple-device, raid1 data/metadata profile, UEFI Secure Boot. A drive 
>> dies, and the system needs to be rebooted before a rebuild occurs.
>> 
>> This works today with /boot on raid1/10 Btrfs. Yes, I manually have to add a 
>> degraded mount option as this isn't automatically done by Btrfs yet. But the 
>> GRUB boot part works. Even degraded, the path to grub.cfg is valid. And the 
>> file system itself keeps multiple copies so there's no work keeping it 
>> current.
>> 
>> With Secure Boot it's a problem. The signed grubx64.efi has a fixed prefix 
>> location, valid on any computer, to search for grub.cfg, which is on the 
>> ESP. So now we need to have multiple copies of grub.cfg, somehow synced, on 
>> each ESP. Or another solution. If we had a Btrfs subvolumetypeGUID, 
>> analogous to the GPT partitiontypeGUID, and specify that as the baked in 
>> partuuid for a signed grubx64.efi to search for /boot/grub/grub.cfg.
> 
> Isn't search --fs-uuid sufficient for this task? Or I'm afraid that I
> didn't understand your problem enough.
> 
>  cat <<EOF > /boot/efi/EFI/<DISTRIBUTION>/grub.cfg
>  search --fs-uuid --set=root `grub-probe --target=fs_uuid /boot/grub`
>  set prefix=(\$root)/boot/grub
>  configfile \$prefix/grub.cfg
>  EOF
> 
>  grub-mkconfig -o /boot/grub/grub.cfg
> 
> This way we can avoid calling "grub-mkconfig -o
> /boot/efi/EFI/<DISTRIBUTION>/grub.cfg" and hopefully can get rid of
> the problem you have.

Yes I think that will work. Do you agree that a GUI installer permitting, e.g. 
/boot on Btrfs raid1, should implement this? Or should it be a future feature 
of grub-mkconfig to figure out, that two grub.cfgs are needed?


Chris Murphy