Re: [Qemu-devel] [PATCH 01/15] s390x: fix virtio feature bitmap

[Alexander Graf]

On 10.04.2011, at 22:06, Aurelien Jarno wrote:

> On Sun, Apr 10, 2011 at 09:26:26PM +0200, Alexander Graf wrote:
>> 
>> On 10.04.2011, at 21:25, Aurelien Jarno wrote:
>> 
>>> On Mon, Apr 04, 2011 at 04:32:10PM +0200, Alexander Graf wrote:
>>>> The feature bitmap in the s390 virtio machine is little endian. To
>>>> address for that, we need to bswap the values after reading them out.
>>>> 
>>>> Signed-off-by: Alexander Graf <address@hidden>
>>>> ---
>>>> hw/s390-virtio-bus.c |    4 ++--
>>>> 1 files changed, 2 insertions(+), 2 deletions(-)
>>>> 
>>>> diff --git a/hw/s390-virtio-bus.c b/hw/s390-virtio-bus.c
>>>> index 58af164..60e0135 100644
>>>> --- a/hw/s390-virtio-bus.c
>>>> +++ b/hw/s390-virtio-bus.c
>>>> @@ -223,7 +223,7 @@ void s390_virtio_device_sync(VirtIOS390Device *dev)
>>>>    cur_offs += num_vq * VIRTIO_VQCONFIG_LEN;
>>>> 
>>>>    /* Sync feature bitmap */
>>>> -    stl_phys(cur_offs, dev->host_features);
>>>> +    stl_phys(cur_offs, bswap32(dev->host_features));
>>> 
>>> Is bswap32 correct here for both big and little endian guests? I don't
>>> really understand the reason why a bswap is needed here, especially
>>> given that AFAIK this code was already used when using KVM.
>> 
>> This is target specific code. The s390-virtio-bus is s390 specific. And yes, 
>> the code was also broken with KVM. That's how I first found it actually.
>> 
> 
> So in short, s390-virtio-bus is specific to big endian machines, but is
> using little endian? That looks a bit weired. I guess it's probably to
> late to change the specification anyway...

Yeah. In fact, it mixes big and little endian pieces. But it really is too late 
to change the (unwritten) spec, unfortunately.


Alex