Re: [Qemu-devel] [PATCH 1/8] virtio_get_byteswap: function for endian-ambivalent targets using virtio.

[Peter Maydell]

On 12 August 2013 10:56, Benjamin Herrenschmidt
<address@hidden> wrote:
> On Mon, 2013-08-12 at 10:52 +0100, Peter Maydell wrote:
>> On 12 August 2013 10:50, Benjamin Herrenschmidt
>> <address@hidden> wrote:
>> > I must be confused ... you mentioned in a previous discussion around
>> > endianness that on some ARM cores at least, when changing the OS
>> > endianness, you had to configure a different lane swapping in the bridge
>> > to the the IO devices (AXI ?)
>>
>> No, that's just the implementation -- the bit in the control
>> register is effectively controlling whether there is byte lane
>> swapping in the part of the CPU which is the data path between
>> it and its bus to the outside world.
>
> I find it amazing that an OS can touch that without hitting the
> hypervisor :-)

It's no different to having a userspace process able to have
a different setting from the OS, really. (There is an equivalent
bit in another register that controls what endianness we
use if we trap to hyp mode.)

> Anyway, ok, we do need to poll from virtio then, but we
> probably need to cache as well, no ?
>
> When do you sample it in qemu ?

It's a bit theoretical at the moment since QEMU's ARM code
kind of assumes little endian. I would expect that at the
point when virtio was in an MMIO callback the CPUState
struct would have been updated via the usual sync process in
kvm_arch_get_registers().

-- PMM