Re: [Qemu-devel] [0/8] Clean up PCI code to allow for multiple root buses

[Michael S. Tsirkin]

On Thu, May 09, 2013 at 10:31:04AM +1000, David Gibson wrote:
> The current PCI subsystem has kind of half-hearted support for
> multiple independent root buses - aka PCI domains - in the form of the
> PCIHostBus structure and its domain field.  However, it doesn't quite
> work because pci_host_bus_register() is always called with a domain of
> 0.
> 
> Worse, though, the whole concept of numbered domains isn't general
> enough.  Many platforms can have independent root buses (usually on
> wholly independent host bridges), but only x86 gives them a
> hardware-significant domain number, essentially as a hack to allow all
> the separate config spaces to be accessed via the same IO ports.
> Linux guests on other platforms will show domain numbers in lspci, but
> these are purely guest assigned, so qemu won't know about them.
> 
> This patch series, therefore, removes the broken-as-is domain concept
> from qemu, and replaces it with a different way of handling multiple
> root buses, based on a host bridge class method to provide a
> identifier for the root bus.  This hook is designed in such a way as
> to allow a single bridge object to support mutiple root buses with
> future changes, which will allow future implementations of x86 north
> bridges with multiple domains to be supported correctly, and in way
> that matches the existing practice for all external interfaces.

I agree with the direction this patchset takes.
I sent some comments on the individual patches separately.

Thanks!


-- 
MST