Re: [PATCH] VT-d: Fix iommu map page for mmio pages

[Anthony Liguori <anthony@codemonkey.ws>]

Muli Ben-Yehuda wrote:
> On Thu, Sep 25, 2008 at 05:45:30PM +0300, Avi Kivity wrote:
>   
>> Han, Weidong wrote:
>>     
>>> Is it possible DMA into an mmio page? 
>>>       
>> I don't see why not.
>>     
>
> Two reasons. First it makes no sense. MMIO pages don't have RAM
> backing them, they have another device's register window. So the
> effect of DMA'ing into an MMIO page would be for one device to DMA
> into the register window of another device, which sounds to me insane.
>   

MMIO isn't just a register window.  It may be an on-device buffer.  For 
instance, all packets are stored in a buffer on the ne2k that's mapped 
via mmio.  It would seem entirely reasonable to me to program an IDE 
driver to DMA directly into the devices packet buffer.

> Second, and more importantly, I've seen systems where doing the above
> caused a nice, immediate, reboot. So I think that unless someone comes
> with a valid scenario where we need to support it or something breaks,
> we'd better err on the side of caution and not map pages that should
> not be DMA targets.
>   

Xen maps the MMIO pages into the VT-d table.  The system you were using 
could have just been busted.  I think the burden is to prove that this 
is illegal (via the architecture specification).

Regards,

Anthony Liguori

> Cheers,
> Muli
>