Re: AMD 8131 and MSI quirk

[Greg KH <greg@kroah.com>]

On Sat, Oct 22, 2005 at 05:32:20PM -0600, Matthew Wilcox wrote:
> On Sat, Oct 22, 2005 at 03:14:34PM -0700, Roland Dreier wrote:
> > The current quirk_amd_8131_ioapic() function sets a global
> > pci_msi_quirk flag, which disables MSI/MSI-X for all devices in the
> > system.  This is safe but suboptimal, because there may be devices on
> > other buses not related to the AMD 8131 bridge, for which MSI would
> > work fine.  As an example, see the end of this email for a lspci -t
> > from a real Opteron system that has PCI-X buses coming from an AMD
> > 8131 and PCI Express buses coming from an Nforce4 bridge -- MSI works
> > fine for the Mellanox InfiniBand adapter on the PCIe bus, if we allow
> > it to be enabled.
> > 
> > I guess what we really should be doing is setting the dev->no_msi flag
> > for all devices below the AMD 8131 PCI-X bridge rather than turning
> > off MSI globally.  Of course this is somewhat tricky, since a device
> > could be hotplugged onto a bus below the AMD 8131.  Greg, any thoughts
> > about the proper way to use the driver model infrastructure to handle
> > this?
> 
> Perhaps the right thing to do is to change pad2 (in struct pci_bus) to
> bus_flags and make bit 0 PCI_BRIDGE_FLAGS_NO_MSI ?

Yeah, I can't think of any way to use the device tree to do this, so
this sounds as good a way as any.

thanks,

greg k-h