On Tue, Nov 27, 2012 at 05:07:31PM +1100, David Gibson wrote: > Hi Michael, Alex, > > This patch represents a compromise I hope will be acceptable after the > long thread discussing handling of multiple PCI host bridges on the > pseries machine. Patch 1/1 is just a preliminary enforcing uniqueness > of LIOBNs in the IOMMU code. > > Patch 2/2 is the meat. It allows either explicit configuration of all > the properties, or the user can just set an abstract index which will > generate sensible and probably-unique values for all the rest. > > With these patches I was able to do something like: > qemu-system-ppc64 -M pseries -m 1024 -nographic \ > fc17-root.qcow2 -net none -device nec-usb-xhci -device \ > spapr-pci-host-bridge,index=1 -device e1000,netdev=fred \ > -netdev user,id=fred > > I was able to see both the PCI domains in the guest, and use the NIC > on the secondary domain. > > There are still some gotches with multiple domains though. The domain > value in PCIHostBus is still always initialized to 0, and there are > other places in the PCI core where handling of multiple domains is > essentially stubbed out. > > Michael, any thoughts on what to do about that? I could fix up the > PCI code so that domain is actually set and used. But I think the > whole notion of domain numbers is kind of bogus on the qemu side: > since PCI domains are completely independent from each other, it's > only platform convention which determines what the domain numbers are. > On platforms that don't have a strong convention, the guest will > number them itself and we have no way of knowing that. So it seems to > me that the PCI code should instead of domain numbers just use the > device ID, or the bus name or some qemu side symbolic name. For > platforms that do have a numbering convention those names can be > derived from the domain numbers, but it also works for platforms that > don't.
I agree with this last statement: using bus numbers and domain numbers is not a good idea. We mostly need to support domain 0 to mean "default" for backwards compatibility. I'll review the patchset shortly. -- MST
