Re: Big Endian fix patch (was: Re: Couple more fixes for Linux raid metadata 1.x support)

[Lennart Sorensen]

On Mon, Jul 26, 2010 at 09:00:53PM -0400, Doug Nazar wrote:
> I'd worried about endianess while testing. I've been spending the last 2  
> days trying to setup a QEMU powerpc64 image to run some regression  
> testing but it's taking forever (combination of old P4 computer and QEMU  
> giving me random errors from the cdrom drive).

It is faster and simpler to use an iso image with qemu rather than the
real cdrom drive.  Seems much more reliable that way.

> I'll run it through the grinder in a bit.

I have an IBM p520 here (dual core power6+ 4.2GHz).  Nothing takes long
here. :)

So with my patch I am booting of the md raid1 very successfully (other
than manually having to fix grub-mkimage run since grub-install doesn't
understand the machine yet.  I am working on fixing that now.)

Quite simply, _everything_ that accesses the superblock structure
numbers has to use endianess conversion since 1.x md superblocks are
always little endian.  0.9 is host endian all the time and hence simpler
(but also not portable).

And the checksum needs folding according to the kernel before comparing
(which I added).  Otherwise different architectures may have different
results.

-- 
Len Sorensen