Re: [RFC][PATCH] unique sigs in the msdos label

[Matt Domsch]

I seem not to be subscribed to bug-parted right now, so please cc: me on 
this thread.  I'm also out on vacation, so I'll respond when time allows.

On Sat, 3 Jul 2004, Andrew Clausen wrote:

> I'll answer for Matt :)
> 
> On Mon, Jun 28, 2004 at 03:46:21PM -0700, John Gilmore wrote:
> > I think writing 'unique signatures into disk labels' is a terrible idea.

I wasn't so keen on it myself I'll admit.  Which is why I've put a lot
of effort into using the "BIOS Enhanced Disk Drive Services 3.0"
specification, which EFI relies on.  Problem is, I can count the
number of correct BIOS implementations of EDD 3.0 on one hand.  I
can't count the number of incorrect or didn't-bother implementations,
though I've tried.  See my listing at
http://linux.dell.com/projects.shtml.

> > Intel is working on a Legacy-BIOS-free spec.
> 
> GPT, yeah.  (Parted and Linux even support it :)
> There are plenty of legacy users left though.

yep.  Andrew and I wrote the GPT stuff for Parted several years ago.

> > It should provide a way to determine exactly what device is the
> > boot device -- without this hocus pocus about storing different
> > magic numbers in every partition table, retrieving them via
> > different routes to compare and contrast.  By getting involved in
> > the writing of this spec

EDD is exactly this spec.  www.t13.org's 1572D project is the latest.
I believe this is what you're referring to, yes?

It's not the writing of the spec that's problematic (though there are
a few problems with the spec, which I've described to the T13
committee chair, they're corner cases right now).  It's BIOS
implementations that are problematic, and I guarantee, there are more
(closed source) implementations than there are people who care to fix them.

> >, we should be able to put a stake through
> > the heart of this issue once and for all, in new hardware.  Like
> > the concept of cylinders, 'BIOS drive numbers' should hsve been
> > d-e-a-d two decades ago.

That was my hope when I wrote EDD for Linux...  it's a good theory,
but not yet practical.

> > As far as I know, nobody cares what device numbers the BIOS assigned
> > to devices other than the boot disk.

Except the one thing that really cares is the OS installer.  Red Hat's
Anaconda, and SuSE's YAST, for example.


> > The main question is, "What
> > disk and partition did I, this kernel/bootstrapper/etc, get read in from?"

Take a perfectly new system, with zeros across all the disks, and yes,
you've got multiple disks attached to multiple PCI devices in the
system.  Which disk should your /boot, / and MBR-based boot loader
(grub, lilo) go, such that after install, the system will boot
properly to that disk?

> With the current BIOS setup, we need to have unique hard disks to
> answer this question.
> 
> We can ask the BIOS which disk is the boot disk, but it will give
> an identifier that "BIOS-only".  i.e. we can't match it up to an IDE
> controller or whatever.  (Reason: the BIOS can do funky remapping
> tricks)  What we can do is ask the BIOS to read sectors from this
> boot device.  If they are unique, we can figure out which disk it is.
> 
> Matt: is this what you had in mind?

Yep.  I wish we didn't have to write sigs to the disk.  I really do.
But it's unavoidable until EDD3.0 works everywhere, which won't happen
any time soon.
 

-- 
Matt Domsch
Sr. Software Engineer, Lead Engineer
Dell Linux Solutions linux.dell.com & www.dell.com/linux
Linux on Dell mailing lists @ http://lists.us.dell.com