iommu external module

iommu external module

[Avi Kivity]

When userspace support for device assignment is merged, we will finally 
have full support for device assignment.  Unfortunately, many users 
won't be able to test or use device assignment, since very few actually 
run development kernels.  The first stable kernel with enough iommu 
support for kvm will be 2.6.28, released in about three months, and 
distros are even further away.

We can fix this fairly simply by having an external module for the 
iommus, much like kvm itself.  I don't think it should be too difficult, 
and it will provide a lot of testing to us, and important functionality 
for our users.

Anyone willing to pick this up?

-- 
error compiling committee.c: too many arguments to function

Re: iommu external module

[Muli Ben-Yehuda]

On Thu, Oct 02, 2008 at 03:45:33PM +0300, Avi Kivity wrote:

> We can fix this fairly simply by having an external module for the
> iommus, much like kvm itself.  I don't think it should be too
> difficult, and it will provide a lot of testing to us, and important
> functionality for our users.

The problem is that if we want to enable DMA translation for a given
device, we have to do it before it has any outstanding DMAs. As long
as you have a device that you want to enable translation for as
built-in, you'll need the IOMMU built-in as well, or will need to
reset the device when you load the IOMMU. That's why all IOMMUs are
currently built-in.

Cheers,
Muli
-- 
The First Workshop on I/O Virtualization (WIOV '08)
Dec 2008, San Diego, CA, http://www.usenix.org/wiov08/
                       <->
SYSTOR 2009---The Israeli Experimental Systems Conference
http://www.haifa.il.ibm.com/conferences/systor2009/

Re: iommu external module

[Avi Kivity]

Muli Ben-Yehuda wrote:
> On Thu, Oct 02, 2008 at 03:45:33PM +0300, Avi Kivity wrote:
>
>   
>> We can fix this fairly simply by having an external module for the
>> iommus, much like kvm itself.  I don't think it should be too
>> difficult, and it will provide a lot of testing to us, and important
>> functionality for our users.
>>     
>
> The problem is that if we want to enable DMA translation for a given
> device, we have to do it before it has any outstanding DMAs. As long
> as you have a device that you want to enable translation for as
> built-in, you'll need the IOMMU built-in as well, or will need to
> reset the device when you load the IOMMU. That's why all IOMMUs are
> currently built-in.
>   

This is for using with kvm as an external module, so there is no 
question of a built in device.  It's purely for older kernels; I'm not 
calling for modular iommus.

-- 
error compiling committee.c: too many arguments to function