Re: [patch] set prefix on PPC

[Hollis Blanchard]

On Feb 13, 2005, at 12:35 PM, Marco Gerards wrote:

> Hollis Blanchard <address@hidden> writes:
>
> > This code sets the GRUB "prefix" environment variable from the OF
> > /chosen/bootpath property. It must therefore translate between the two
> > syntaxes.
>
> Personally I don't like setting prefix to the path grubof was loaded
> from.  Mainly because of the way how things work on the macintosh.  It
> has a special HFS[+] boot partition (at least my installation, but I
> think this is the common practice).
>
> I don't like installing all modules and the configuration file on this
> partition.  There were some problems with writing to this partition
> IIRC and it is not comfortable to use.  It should not be touched by a
> user, I just can't remember why.  I will look that up.

Please let me know if you find any actual data. I have heard the ybin  
author rant about this for years (see  
http://lists.penguinppc.org/yaboot-users/2002/yaboot-users-200210/ 
msg00008.html as an example), but he doesn't like to substantiate his  
claims.

I believe there were two problems with keeping /boot on an HFS  
partition:
1) The Linux HFS driver wasn't overly reliable. That's unfortunate if  
you're writing a kernel there, or editing grub.cfg. For that reason,  
/boot is typically kept as a Linux-native filesystem, and to install a  
kernel ybin invokes hfsutils to do the copying to the "bootstrap"  
partition. (Note the similarity in names "boot" and "bootstrap" causes  
confusion.)
2) HFS has a concept called "blessed"; basically the firmware can  
automatically find and boot blessed executables without needing to  
specify a file name or even partition. Some versions of Mac OS  
allegedly will unbless a non-Mac OS blessed file. For that reason, ybin  
prefers the HFS "bootstrap" partition be of a type that Mac OS will not  
mount (and so will not alter).

However, please do not mistake "this is how ybin does it" for "this is  
the only way to do it". I have never seen these problems (though I  
personally ran an HFS /boot partition that Mac OS mounted), and I have  
never seen data as to what exact software versions or scenarios cause a  
problem. The user in the URL above claims no problems either, and I  
haven't heard of other users with these problems.

> What I would prefer is a way to encode the prefix into the grubof
> binary.  The only disadvantage is that it might be hard to detect this
> install time.

If possible, placing grub.cfg next to grubof will be the easiest  
solution. That way, no matter how or where you boot GRUB (even from  
rescue CD or whatever), you will always know what grub.cfg it will  
load.

Example: you're making a distribution installer and need to boot from  
CD. What prefix do you embed in GRUB? There is no guarantee that every  
machine out there has a "cd" devalias, nor would you want to require  
users to create that in OF before booting. Simply looking adjacent to  
the grubof binary is a much better solution.

> What I would prefer is this:
>
> 1)  Set prefix to whatever was set in the binary.

This would involve more post-compilation ELF manipulation. To  
accomplish that, if we do decide to go that route, I think it would be  
worth converting the existing module segment into a more  
general-purpose segment, containing all the added arguments.

Remember, right now we load modules at a fixed address. At runtime,  
GRUB finds at that address:
1. the size of the modules in memory
2. the module binaries

We can change this to be more general-purpose, e.g. a chain or  
"parameter" structures that could contain the module binaries, the size  
of modules, the prefix, etc.

> 2)  If nothing was set, use `bootpath'.

So it's not such a bad idea after all huh? :)

> 3)  Read the arguments and if a prefix was passed to GRUB, use it to
>     override the prefix.
>
> For #3 I wrote a patch, but I forgot to send it to grub-devel... It's
> included with this email.  As far as I am concerned #1 is what is
> really important for us.
>
> 2005-02-13  Marco Gerards  <address@hidden>
>
>         * kern/powerpc/ieee1275/init.c (grub_machine_init): Initialize the
>         prefix env variable with the bootargs when it has a valid value.

I don't like that this assumes bootargs only contains the prefix. For  
example, if we ever get our fancy debugging support, it would be very  
nice to boot grub with "debug=all" as an argument.

>         * loader/powerpc/ieee1275/linux.c (grub_rescue_cmd_linux): Don't
>         initialize `init_addr' yet, set it to 0 instead.

Confused... how is this related to bootargs and prefix?

-Hollis