Hi Neil,

No.  I would not accept patches.
Not because I fundamentally disagree with autostart, but because it
wouldn't work.  I suggest you try it.

Maybe I didn't explain very clearly what my goal is...allow me to
clarify:

As it is now, if you compile md as a module and happen to have a
partition tagged as RAID, you'll panic the kernel. There's code in
fs/partitions (and in 2.5, in init/do_mounts.c) that is calling into
md....yechhh. What I'd like to do is make a clean separation between the
boot up and partitioning code and the md driver, so md can be compiled
as a module. (This would also allow other volume managers to use the
partition tagging to identify devices, if they wanted.) 

quoted

Or is there a better way to do this?

Yes.
initrd/initfs is "the right way".
As far as I know, it *does* work.(*)  You do not need md linked into the
kernel.
You have md.o and whatever personalities you want in the initrd image,
then in the initrd script (linuxrc I think) you load the modules, and
use mdadm to assemble the raid arrays.
You then mount the real root and off you go.

Right, sorry, I wasn't clear. This is basically what I meant. All I want
to do is allow autostart_arrays() to be called when md is compiled as a
module. Then, let the initrd/linuxrc (later initramfs) call
autostart_arrays() (ioctl("/dev/md0", RAID_AUTORUN) is all, really).

I've hacked on 2.5.63. I'll attach the patch, it's fairly short...and
might give you a better idea what I'm proposing...

Any comments would be appreciated.

Thanks,
Paul