PV PCI passthrough - summary/status

PV PCI passthrough - summary/status

[Konrad Rzeszutek Wilk]

There has been a sudden flurry of folks using the PCI pass-through and I thought
this would be a good time to give an update on its status in the PV OPS side.

Please keep in mind this is for PV guests, not HVM (well, kind of - there is one
overlap - BAR page alignment).

There are four outstanding issues:
  - MSI support (No API for pci-back to pass the domain ID to Xen's Dom0 MSI support:
    xen_setup_msi_irq).  One idea I am slowly prototyping is for xen_create_msi_irq
    to enumerate the XenStore keys and see if the PCI device in question also
    residies in /local/domain/0/pci/X/Y/ and extract the domain ID from that.

    The beauty of this is that it can also be used by the Xen-SWIOTLB to determine
    whether it should turn itself on for DomU guests.

  - PCI front support. Haven't started on this.

  - BAR page alignment. Looking to see if Xen/xc.c can handle mapping of
    a offset within a page. This way this requirment (need a page-aligned BAR)
    can be eliminated. Or if this is a no-go see if there is a security risk
    in passing the whole page to the guest. Lastly if none of this works, revisit
    the 2.6.18 'reassign_resources' patch that sets the BARs to page-aligned addresses.

  - Bug #1340: The Function Level Reset reset requires that PCI devices behind a bridge
    be owned by the same guest. Based on the comments, without the FLR capability in 3.2, the
    PCI devices worked fine.  We could provie a 'flr_inhibit' flag which would revert
    back to 3.2 and not do an FLR?

Any suggestions, etc., are much welcome.