The VFIO pci device assignment driver will be coming with the release of linux kernel 3.6. VFIO , or virtual function IO as it was originally called, will bring improved IO performance of pci devices with KVM virtual machines. The project was spawned out of the KVM project and has been in development since about 2010 by developer Alex Williamson of Red Hat. This driver is not architecture specific but it does require systems with IOMMU such as those provided by intel vt-d or AMD's AMD-vi. In essence this provides userspace direct access to pci devices in a secure manner and since KVM systems emulation runs out of userspace, KVM will benefit from this driver.
Patches for a userspace driver have already been submitted to qemu 1.2 branch for x86 systems scheduled for September 2012, less than a month from now. The qemu patches only make use of pci devices that use MSI and MSI-x ie there's no support for PCI INTx at this time. Support for INTx is expected to come later. No patches for this feature have been added to libvirt yet but no doubt this feature will move up the stack to libvirt followed by tools such as virt-manger and oVirt as soon as the interface has been stabilized in qemu.
KVM leverages the linux kernel heavily but this is an example where the KVM project is providing some useful value to the kernel at large since there are other userspace applications that will benefit from this driver.