Re: [RFC PATCH v2 08/11] powerpc: gamecube/wii: early debugging using usbgecko

[Benjamin Herrenschmidt <benh@kernel.crashing.org>]

On Mon, 2009-11-30 at 06:50 +0100, Albert Herranz wrote:
> Benjamin Herrenschmidt wrote:
> > On Sat, 2009-11-28 at 21:43 +0100, Albert Herranz wrote:
> >> +        * Prepare again the same BAT for MMU_init.
> >> +        * This allows udbg I/O to continue working after the MMU is
> >> +        * turned on for real.
> >> +        *
> >> +        * We are assuming here that exi_io_base is identity mapped.
> >> +        */
> >> +       addr = ((unsigned long)exi_io_base) & 0xffff0000;
> >> +       setbat(1, addr, addr, 128*1024, PAGE_KERNEL_NCG); 
> > 
> > How do you prevent that from overlapping otherwise valid kernel
> > mappings ?
> > 
> 
> ug_udbg_init() is called from ppc_md.init_early.
> It doesn't overlap any valid kernel mappings because exi_io_base is
> hardcoded to an i/o region not used yet by the kernel.

But that doesn't allocate virtual space does it ?

> See udbg_early_grab_exi_io_base().

Yeah, I see that:

+#if defined(CONFIG_GAMECUBE)
+       return (void __iomem *)0x0c006800;
+#elif defined(CONFIG_WII)
+       return (void __iomem *)0x0d006800;
+#else

So you'll have BATs floating over user addresses ? That sounds fishy :-)

It should be trivial to just create a fixmap entry instead. That will
give you a virtual address that is known at compile time (so you can
use it from head_32.S as well for setting up your BAT).

Actually you probably need more than one entry in there since it needs
to be big enough to cover a BAT min size and be aligned, but it's not
-that- hard to do (I think x86 does similar tricks in their fixmap
iirc) 

> The setbat just prepares again, exactly in the same way, the same BAT that we got
> setup by setup_usbgecko_bat in head_32.S.

 .../...

> > You need to allocate the virtual space. For a debug thing like that, you
> > could use the fixmap. In fact, I think we should create a fixmap entry
> > or two always available for use by early debug.
> > 
> 
> Or give us back ppc_md.setup_io_mappings :)

Not happening :-) Or if you get one, it will allocate virtual addresses
and so you cannot rely on an identity mapping.