[PATCH v2 3/4] arm64: dts: Add mediatek MT8173 SoC and evaluation board dts and Makefile

[Marc Zyngier]

Hi Eddie,

On 2014-12-23 08:02, Eddie Huang wrote:
> Hi,
>
> On Mon, 2014-12-22 at 09:07 +0000, Marc Zyngier wrote:
>> On 20/12/14 20:07, Arnd Bergmann wrote:
>> > On Wednesday 17 December 2014 15:01:29 Marc Zyngier wrote:
>> >>
>> >> Indeed, as described in the documentation:
>> >> 
>> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0471b/CHDIFAEE.html
>> >>
>> >> Also it is worth noticing that given how GICV is placed, it will 
>> never
>> >> work with 64K pages and virtualization. Pretty sad.
>> >
>> > Does this mean no VGIC support on this platform so you have to 
>> emulate it
>> > in order to run virtual machines with 64K pages, or does it mean 
>> that
>> > it's impossible to use that way because you can't emulate it?
>>
>> As Peter said, this is not a configuration we're willing to support:
>> - we don't have a API to tell userspace emulation about interrupts
>> generated by the generic timers
>> - we could move the whole GIC emulation into the kernel (at the 
>> moment,
>> only the distributor is there), but that would be a complete 
>> nightmare
>>
>> It really looks like a case of "let's drop a bunch of 64bit cores 
>> into
>> an existing SoC". Shame people can't read integration guidelines...
>>
>> 	M.
>
> MT8173 use GIC-400. We check GIC-400 TRM that VGIC address is not 
> 64KB
> alignment. but GIC-500 (GICv3) VGIC base address is. We also check
> 3.19-rc1 arm64 device tree, amd-seattle-soc.dtsi VGIC is 64KB 
> alignment,
> but arm juno.dts is not, they are both GIC-400. So we are a little
> confused, and still try to figure out what is the correct address.

I'm afraid you're missing the point.

It is not a matter of alignment, but a matter of having put the GICV
region in the same 64K page as the rest of the GIC. Which means that
if the hypervisor (KVM or Xen) wants to map the GICV region inside the
guest, it will also be forced to give it access to GICC, GICD, and 
GICH.

Goodbye isolation, see your hypervisor going down at the first buggy 
guest,
not to mention all kind of other nice hacks...

I could understand the need to tightly pack everything on a 32bit 
system,
but doing so on a 64bit system without consideration of the page size
is a bit of a bummer.

> Neverthless, MT8173 SoC already use 0x10224000 and 0x10225000 as VGIC
> base address.

Then its fate is sealed virtualization wise.

Thanks,

         M.
-- 
Fast, cheap, reliable. Pick two.