Re: [OpenBIOS] [Qemu-ppc] [PATCH 1/7] pci: add MPC105 PCI host bridge emulation

[Blue Swirl]

On Tue, May 7, 2013 at 5:48 AM, Hervé Poussineau  wrote:
> Andreas Färber a écrit :
> > Am 06.05.2013 22:57, schrieb Hervé Poussineau:
> > > Alexander Graf a écrit :
> > > > On 05/03/2013 07:57 AM, Hervé Poussineau wrote:
> > > > > Alexander Graf a écrit :
> > > > > > Am 02.05.2013 um 22:08 schrieb Hervé Poussineau
> > > > > > :
> > > > > >
> > > > > > > Non-contiguous I/O is not implemented.
> > > > > > >
> > > > > > > There is also somewhere a bug in the memory controller, which means
> > > > > > > that some real firmwares may not detect the correct amount of memory.
> > > > > > > This can be bypassed by adding '-m 1G' on the command line.
> > > > > > >
> > > > > > > Add x-auto-conf property, to automatically configure the memory
> > > > > > > controller at startup. This will be required by OpenBIOS, which
> > > > > > > doesn't know how to do it.
> > > > > >
> > > > > > Why not teach it? I'd prefer to see that logic in firmware.
> > > > >
> > > > > Me too, but I'm not confident enough in my capabilities to do it.
> > > >
> > > > Huh? Why not? Most of the device initialization code in OpenBIOS
> > > > happens in C, so you don't even have to touch Forth code :).
> > > > > Autoconfiguration is only in one place of the code, so I think it can
> > > > > be removed easily once OpenBIOS has this logic.
> > > >
> > > > I'd prefer if we could come up with a clean model from the start. It
> > > > really shouldn't be hard at all.
> > >
> > > I thought that for all other usages of OpenBIOS in QEMU, RAM was
> > > supposed to be available as soon as machine was powered on.
> > >
> > > However, I checked OpenBIOS code:
> > > One of the first things done in arch/ppc/qemu/start.S is to copy the
> > > exception vectors. So, I should add code before it to detect memory
> > > controller, detect ram size and configure memory controller?
> >
> > No. Why? QEMU does not depend on the memory controller being
> > initialized, only the OS might expect some registers to be filled in. So
> > you should look at or add the MPC105 PHB initialization hook in
> > OpenBIOS' PCI code, long after the memory is set up.
>
> OpenBIOS depends of memory being available (at least the first KB/MB) even
> at its very startup, in arch/ppc/qemu/start.S. PCI initialization code comes
> much later.

OpenBIOS for Sparc32 and Sparc64 does not use RAM before RAM size has
been read from fw_cfg. A check for QEMU signature is done and machine
ID is also read before that.
arch/sparc32/entry.S
arch/sparc64/entry.S

It should be easy to change PPC to check the machine ID before accessing RAM.
> At boot, MPC105 datasheet says that memory controller is not configured, ie
> you have to not use RAM in OpenBIOS before PCI initialization.

The memory controller should be initialized much earlier than PCI.
> For other PPC targets (mac99, g3beige) using OpenBIOS, RAM is accessible at
> startup, so that's not a problem for OpenBIOS.
>
> So, no, QEMU does not depend of the memory controller being initialized, but
> OpenBIOS depends of the RAM being accessible ways before PCI initialization.
> I don't speak of the OS which might (or might not) expect some registers to
> be filled in.
> x-auto-conf property correctly sets some registers, so that memory is
> available at startup (like on mac99, g3beige emulations).
>
> Hervé