[RFC 0/2] VFIO: Add virtual MSI doorbell support.

[Alex Williamson]

On Tue, 2015-07-28 at 17:55 +0100, Marc Zyngier wrote:
> Hi Alex,
> 
> On 28/07/15 17:21, Alex Williamson wrote:
> > On Fri, 2015-07-24 at 14:33 +0530, Pranavkumar Sawargaonkar wrote:
> >> In current VFIO MSI/MSI-X implementation, linux host kernel
> >> allocates MSI/MSI-X vectors when userspace requests through vfio ioctls.
> >> Vfio creates irqfd mappings to notify MSI/MSI-X interrupts
> >> to the userspace when raised.
> >> Guest OS will see emulated MSI/MSI-X controller and receives an interrupt
> >> when kernel notifies the same via irqfd.
> >>
> >> Host kernel allocates MSI/MSI-X using standard linux routines
> >> like pci_enable_msix_range() and pci_enable_msi_range(). 
> >> These routines along with requset_irq() in host kernel sets up 
> >> MSI/MSI-X vectors with Physical MSI/MSI-X addresses provided by
> >> interrupt controller driver in host kernel.
> >>
> >> This means when a device is assigned with the guest OS, MSI/MSI-X addresses
> >> present in PCIe EP are the PAs programmed by the host linux kernel.
> >>
> >> In x86 MSI/MSI-X physical address range is reserved and iommu is aware
> >> about these addreses and transalation is bypassed for these address range.
> >>
> >> Unlike x86, ARM/ARM64 does not reserve MSI/MSI-X Physical address range and
> >> all the transactions including MSI go through iommu/smmu without bypass.
> >> This requires extending current vfio MSI layer with additional functionality
> >> for ARM/ARM64 by
> >> 1. Programing IOVA (referred as a MSI virtual doorbell address)
> >>    in device's MSI vector as a MSI address.
> >>    This IOVA will be provided by the userspace based on the
> >>    MSI/MSI-X addresses reserved for the guest.
> >> 2. Create an IOMMU mapping between this IOVA and
> >>    Physical address (PA) assigned to the MSI vector.
> >>
> >> This RFC is proposing a solution for MSI/MSI-X passthrough for ARM/ARM64.
> > 
> > 
> > Hi Pranavkumar,
> > 
> > Freescale has the same, or very similar, need, so any solution in this
> > space will need to work for both ARM and powerpc.  I'm not a big fan of
> > this approach as it seems to require the user to configure MSI/X via
> > ioctl and then call a separate ioctl mapping the doorbells.  That's more
> > code for the user, more code to get wrong and potentially a gap between
> > configuring MSI/X and enabling mappings where we could see IOMMU faults.
> > 
> > If we know that doorbell mappings are required, why can't we set aside a
> > bank of IOVA space and have them mapped automatically as MSI/X is being
> > configured?  Then the user's need for special knowledge and handling of
> > this case is limited to setup.  The IOVA space will be mapped and used
> > as needed, we only need the user to specify the IOVA space reserved for
> > this.  Thanks,
> 
> I guess my immediate worry is that it seems to impose a fixed mapping
> for all the guests, which would restrict the "shape" of the mappings we
> give to a guest. Or did you intend for that IOVA mapping to be defined
> on a "per userspace instance" basis?

Hi Marc,

Right, I'm not suggesting a fixed mapping imposed on the user, I'm
suggesting the user can set aside a range of IOVA space for VFIO to make
use of for this purpose.  The user would be explicitly defining that
range of reserved IOVA space.

We effectively have your first scenario on x86, where the platform
restriction shapes the VM.  When we have an x86 guest on x86 host, the
guest and host implicit reserved regions align and we don't have any
problems.  However if we wanted to run an ARM64 guest with an assigned
device on an x86 host, we'd need to "shape" the VM to avoid memory
overlapping the implicitly reserved range.

Ideally we could extend the interfaces to support both of these, a
mechanism to discover implicitly reserved ranges and for the user to set
explicitly reserved ranges.  Thanks,

Alex