On 23/02/2017 11:02, Peter Maydell wrote:
> On 23 February 2017 at 08:35, Paolo Bonzini <pbonz...@redhat.com> wrote:
>>
>>
>> On 23/02/2017 05:20, Alexey Kardashevskiy wrote:
>>> First, Paolo is right and ram_device_mem_ops::endianness should be
>>> host-endian which happens to be little in our test case (ppc64le)
>>
>> So you tested a ppc64 BE guest and it works?
>>
>>> Keep things where they are in the VFIO department and just fix
>>> ram_device_mem_ops::endianness?
>>
>> I would fix the ram_device_mem_ops. Either by introducing
>> DEVICE_HOST_ENDIAN(*) or with Yongji's patch.
>>
>> (*) DEVICE_NATIVE_ENDIAN is special cased all over the place
>> because the same device (in a file that's compiled just once)
>> can be either little- or big-endian. DEVICE_HOST_ENDIAN can
>> be a simple #define to either DEVICE_LITTLE_ENDIAN or
>> DEVICE_BIG_ENDIAN, because host endianness is the same for
>> all QEMU binaries. It's literally half a dozen lines of code.
>
> I'm really not convinced we need DEVICE_HOST_ENDIAN. RAM
> areas should be target-endian (you can probably define
> "target endianness" as "the endianness that RAM areas have".)
This is not RAM. This is MMIO, backed by a MMIO area in the host. The
MemoryRegionOps read from the MMIO area (so the data has host
endianness) and do not do any further swap:
data = *(uint16_t *)(mr->ram_block->host + addr);
Here, the dereference is basically the same as ldl_he_p.
If you wanted to make the MemoryRegion use DEVICE_NATIVE_ENDIAN, you'd
need to tswap around the access. Or you can use ldl_le_p and
DEVICE_LITTLE_ENDIAN (this is what Yongji's patch open codes), or
ldl_be_p and DEVICE_BIG_ENDIAN. They are all the same in the end.
Paolo
