Re: Move loader.c out of the kernel

[Yoshinori K. Okuji]

On Wednesday 01 April 2009 23:19:32 Robert Millan wrote:
> On Wed, Apr 01, 2009 at 10:52:26PM +0900, Yoshinori K. Okuji wrote:
> > On Tuesday 31 March 2009 17:56:24 phcoder wrote:
> > > With a new swing in normal.mod splitting I think we should reconsider
> > > this patch. It's useless to keep loader.c in kernel without boot
> > > command. IMO it should be moved either to a perate boot.mod (my
> > > preference) or to minicmd.mod (not a good option IMO)
> >
> > As I said, rescue mode is not quite useful without any loader. So the
> > loader interface should be built into the kernel, and the boot command
> > should be as well naturally.
>
> This presses for more space into core.img, which is highly constrained
> (specially in weird combinations like raid + lvm or crypto in the future).
>
> Why is a loader so important for rescue mode?  If the loader would work, it
> means you can read files, so it should be able to load the rest of modules
> as well.

If some or all modules are trashed, no. Also, the user might be able to read 
other partitions.

> When user is dumped to rescue mode, usually (at least for reports I dealt
> with in debian) it means GRUB has a bug or didn't setup itself properly,
> and the /boot/ directory can't be accessed.  A loader wouldn't help in
> these situations.

And, the rescue mode does not help very much in this case.

> Also, how do you determine which loaders belong in kernel?  There can be
> many specialized loaders like the linux one.  Or we could just put
> multiboot, but the Multiboot loader is quite complex already, and it still
> has room for growing.  Maybe the answer is to write a very simple Multiboot
> loader and put that in kernel?

That's up to the user IMO. For example, suppose this kind of scenario:

- the user has a dual boot environment, say, GNU/Linux and Windows.
- she uses GRUB installed with the GNU/Linux.
- she has made something stupid in the GNU/Linux (e.g. rm -rf /).

In the case of GRUB Legacy, this is the end. It's just unbootable, because 
stage 1.5 may not boot anything besides stage2.

In the case of GRUB 2, if the user has pre-loaded chainload.mod onto the 
core.img, she can still boot the Windows.

Honestly, I am more interested in making it possible to use GRUB for setting 
up a very robust environment. For example:

- the user installs a main OS and a rescue OS into a single machine. The 
latter can be very compact (e.g. puppy, grml, etc.).
- she installs another boot loader into the partition for the rescue OS.
- she pre-loads chainload.mod in core.img.

I think the chainloader is the most realistic, because it is the smallest 
loader, and GRUB does not need to read a filesystem, thus no filesystem 
driver is required for loading an OS.

Every time I maintain a remote system, I feel the necessity of being prepared 
for disasters (e.g. filesystem crashes). Rescue mode is one of the tools 
required for this goal.

Regards,
Okuji