[PATCH v2] arm/arm64: KVM: MMIO support for BE guest

[Marc Zyngier]

On 30/10/13 17:19, Anup Patel wrote:
> On Wed, Oct 30, 2013 at 10:41 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
>> On 30/10/13 17:06, Anup Patel wrote:
>>> On Wed, Oct 30, 2013 at 1:58 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
>>>> On 2013-10-30 01:10, Anup Patel wrote:
>>>>>
>>>>> On Wed, Oct 30, 2013 at 12:19 AM, Marc Zyngier <marc.zyngier at arm.com>
>>>>> wrote:
>>>>>>
>>>>>> Do the necessary byteswap when host and guest have different
>>>>>> views of the universe. Actually, the only case we need to take
>>>>>> care of is when the guest is BE. All the other cases are naturally
>>>>>> handled.
>>>>>
>>>>>
>>>>> You might want to handle the case where we have LE guest on BE host
>>>>> because for ARM64 kernel we might have lot of people interested in
>>>>> running host kernel in BE mode with KVM enabled.
>>>>
>>>>
>>>> What makes you think it is not handled already?
>>>
>>> What I understood here is that you are trying to ensure that MMIO data
>>> passed to/from user space (i.e. QEMU or KVMTOOL) is host endian
>>> using vcpu_data_guest_to_host() and vcpu_data_host_to_guest(). This
>>> makes lot of sense for having all combinations of host and guest endianness.
>>>
>>> If the above is correct then I see an issue in vcpu_data_guest_to_host() and
>>> vcpu_data_host_to_guest() for LE guest on BE host because this patch does
>>> endianness conversion for BE VCPUs only in vcpu_data_guest_to_host() and
>>> vcpu_data_host_to_guest(). If we have LE guest on BE host then these
>>> functions won't do any endianness conversion.
>>
>> And no conversion is exactly what we want. MMIO is always LE.
> 
> Ok got it.
> 
> I think vcpu_data_guest_to_host() and vcpu_data_host_to_guest() have
> little misleading names. Can we name it vcpu_data_guest_to_le() and
> vcpu_data_le_to_guest()?

My view is that encoding the endianness in the name is also confusing.
We're actually going from guest-endianness to "no-endianness".

> Now if we are always passing LE MMIO data to QEMU (even for BE host)
> then this will have to be handled properly for BE QEMU which means
> code flow for LE QEMU and BE QEMU will not be same.

Why so? As long as MMIO is dealt with using the same endianness, it
should behave the exact same, and the code should be the same. At least,
it definitely is for kvmtool, and from what I gathered, the PPC guys
didn't have much issues running LE guests on BE hosts using QEMU.

	M.
-- 
Jazz is not dead. It just smells funny...