bootloader & head.S weirdness

[Benjamin Herrenschmidt]

Hi !

I've almost finished a fist version of my new OF-based bootloader (for
newworld only) and I'm struggling with a few issues:

It appears that depending on where I load the kernel (and other various
manipulation with OF mappings), the kernel will sometimes boot or simply
cause various exceptions in head.S, mostly in the relocation code.

I've tried to figure out what's going on and finally decided (am I wrong
?) that what head.S does is fundamentally broken:

The kernel enters from OF, with the MMU in whatever state OF set it and
the guarantee that at least it's mapped 1:1 with some memory (1Mb) also
mapped just after it so that it can expand itself. This guarantees are
the responsibility of the bootloader.

At this point, it hacks a BAT to map the first 16Mb of memory to
KERNELBASE. That I don't like since it means playing with mapping of
potentially running code with MMU enabled (and you don't know if OF
didn't use a BAT itself for mappping the RAM you are running from, future
OF versions may do). It will then do all sort of weird guesses depending
on the bootloader to know if it must translate or not the source address,
do the copy and flush (which will usually default catch to OF with my
bootloader, depending on where I loaded the kernel, etc...)

Quik tries to load the kernel at 0, which avoids most of those head.S
issues, but the way it does is broken, since OF won't allow to CLAIM
memory from 0 (at least not on my machine), that means potentially
overriding bits of OF (I know OF tends to use high memory, but OF low
memory vectors are still needed, at least on 603 for prom_entry to work).

Also, the bootloader can't turn MMU off and clean TLB & BATs since
prom_entry must have the correct OF mappings in order to work.

Instead of trying to find out the magical address to load the kernel,
potentially breaking with future OF versions, and hacks like that, I
finally decided to give a try at fixing head.S (I'm still using 2.2.x
kernels, I beleive head.S did change with 2.3.x, I'll look into this later).

Basically, what I'm doing is, after bl prom_entry, and before filling the
BAT, I turn the MMU off, clear all bats (in a 601-compatible way unlike
the current kernel source tree), flush TLBs, and then continue to the
code that sets up the BAT and moves down the kernel to 0. I also had to
comment out the code that translates the source address on CHRP.

I'll post a patch later today, but I wanted to know if my way of doing
things is broken or will break CHRP machines or not. Basically, I feel
the "good way" is to shut down the MMU upon exit from prom_entry, and
then play with BATs, copy&flush, etc...

Any comments ?


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/