Granularity of grub-install (was Re: Reproducibility of grub-install)

[Miguel Arruga Vivas]

Hi,

First of all, thank you for your comments.  My answers are inline,
although the bad subject was a main issue of misunderstanding.  I
changed it, as I think it reflects better my current idea.  I'm sorry
for the confusion it certainly caused.

Wed, 23 Oct 2019 11:09:07 +0200
Daniel Kiper <address@hidden> wrote:
> On Mon, Oct 21, 2019 at 04:30:21PM +0200, Miguel Arruga Vivas wrote:
> > Hi, everybody!
> >
> > After taking a deeper look into our (guix's) grub installation
> > procedure, I have the thought that it could be a neat idea to make
> > the boot directory an actual derivation instead something of the
> > global status.  
> 
> I do not understand what you mean by "make the boot directory an
> actual derivation instead something of the global status". Could you
> elaborate more about that?

Currently, the /boot directory is partially a system-wide directory,
even though most of its contents are direct copies of the ones from
grub's installation prefix: /usr/... in most distributions including
BSD world, /{nix,gnu}/store/<hash>-grub-2.04 in Nix/Guix.  Most of its
contents don't depend at all on the actual machine and that can be
shared between machine-dependent boot-sector/efi-variables
"activations".  These contents include all grub modules and message
object files.  On the other hand, the task of the platform-specific
installation cannot be "shared", as it must be performed directly "on
the metal".  And there is a third task between these two: the generation
of the raw image, that depends on the hardware configuration through
the provided configuration file *and* the binary installation.

> (...)
> I am not sure why grub.cfg have to be reproducible. OK, to some extent
> it has but...

Guix doesn't generate grub.cfg through grub-mkconfig, as each kernel is
also associated with the configuration for the init system too--GNU
Shepherd in Guix--, and one kernel may boot up different systems, so it
is generated with Guile scripts.  My understanding is that grub-mkconfig
is a portable tool, not only intended for this case, whose output is
system dependant.  I mixed up concepts here, sorry.

> >
> >   - /boot/grub/grubenv: IIUC, this file must be writable by grub.
> > This should not be on the store, and not sharing the path may be the
> >     main problem right now to implement this.  
> 
> I do not understand this.
>
> > AFAIK, the grubenv problem requires a modification of the grub code
> > if we try to use a different path for this kind-of-modifiable file,
> > so this would require modify grub to being able to lookup for that
> > file somewhere else.  This way the global state can be made
> > explicit.  
> 
> What do you mean by "This way the global state can be made explicit."?

Sorry again.  We do not use load_env nor save_env in our
configuration.  I misunderstood the creation in grub-install.c as a
hidden requirement for it, but I'd just had to read the manual.  That
global state already is explicit enough.  I think now that its creation
should be optional, even though it's created by default as
grub-mkconfig makes use of it.  Should I send a patch for it?

> > The image installation into the device is a separate issue from the
> > binaries installation, that could be separated into two separate
> > binaries, or two steps/flags for grub-install, one for binaries
> > installation into ${boot-directory}/grub and the other one for
> > load.cfg generation and core/boot.img installation.  
> 
> Why do you need to separate this thing into two steps?

I'd like being able to have different grub /boot-like folders, maybe
through unions in the file system (as hard links) of the shared
contents and the physical system dependencies, and activate them
selectively without writing into them any more.  This is intended for
Guix's store[1], but other use cases are possible, in order to roll-back
to a previous generation of the system by only writing
core.img/grub.efi/etc. into platform-dependant locations.

As I said before, I see them now as three steps.  The middle one is
already there with grub-mkimage.  The bios-like installation can be
performed with grub-setup, but EFI systems need grub-install to copy
the image and activate it, the reason behind grub-setup deprecation if I
understand the code correctly.  And the copy of the correct files to
the boot-directory needs no modification at all of grub-install.

> > To grub-devel: I'd be able to send patches for the latter if you
> > think  
> 
> Patches are always welcome...

I'll send just to grub-devel following this mail a patch for a minor
problem I've found reading the code. 

The code for an --only-platform-installation flag (the name I'm using
right now) needs more work but I'll send it as soon as I have
something I've tested on x86 BIOS and x86_64 EFI.  From my
current understanding it would replace completely grub-setup, wouldn't
it?

Happy hacking!
Miguel

[1] https://guix.gnu.org/manual/en/html_node/Features.html