2.6.27, kvm-kmod, and PCI device assignment

2.6.27, kvm-kmod, and PCI device assignment

[Chris Friesen]

Hi,

I've been away a bit but I'm now coming back to this issue.

Is it expected that the kvm-kmod modules will allow PCI device 
assignment (aka VT-d) with older kernels?  If so, then is there an 
expectation for how far back it will be supported?  I haven't seen a 
clear statement of this anywhere, and I think David was interested in 
this as well.

Looking at the instructions at 
"http://www.linux-kvm.org/page/How_to_assign_devices_with_VT-d_in_KVM", 
step 4 is to unbind the device from the host using the "pci_stub" 
functionality.  That code was only added to mainline for 2.6.29-rc1, so 
it seems like extra work will be needed for 2.6.28 and earlier.

Thanks,
Chris


-- 
Chris Friesen
Software Developer
GENBAND
chris.friesen@genband.com
www.genband.com

Re: 2.6.27, kvm-kmod, and PCI device assignment

[Jan Kiszka]

On 2011-09-12 19:56, Chris Friesen wrote:
> 
> Hi,
> 
> I've been away a bit but I'm now coming back to this issue.
> 
> Is it expected that the kvm-kmod modules will allow PCI device
> assignment (aka VT-d) with older kernels?  If so, then is there an
> expectation for how far back it will be supported?  I haven't seen a
> clear statement of this anywhere, and I think David was interested in
> this as well.

In general, it is supposed to work, but there are likely more quirks and
limitation than with a recent kernel. Example: pci_reset_function is not
available for < 2.6.28, thus the assigned device may not work properly
if the guest depends on a reset or after the VM ran through a reset or
<who-knows>. You may also run into bugs/limitations of required kernel
subsystems like the IOMMU layer that were fixed in newer kernels.
kvm-kmod only addresses the kvm modules here.

> 
> Looking at the instructions at
> "http://www.linux-kvm.org/page/How_to_assign_devices_with_VT-d_in_KVM",
> step 4 is to unbind the device from the host using the "pci_stub"
> functionality.  That code was only added to mainline for 2.6.29-rc1, so
> it seems like extra work will be needed for 2.6.28 and earlier.

pci_stub is not strictly required if you manage your system in a way
that all other "interested" drivers are kept away from the assigned device.

Jan

-- 
Siemens AG, Corporate Technology, CT T DE IT 1
Corporate Competence Center Embedded Linux