Re: [patch V2 24/46] PCI: vmd: Mark VMD irqdomain with DOMAIN_BUS_VMD_MSI

[Derrick, Jonathan]

Hi Jason

On Mon, 2020-08-31 at 11:39 -0300, Jason Gunthorpe wrote:
> On Wed, Aug 26, 2020 at 01:16:52PM +0200, Thomas Gleixner wrote:
> > From: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > 
> > Devices on the VMD bus use their own MSI irq domain, but it is not
> > distinguishable from regular PCI/MSI irq domains. This is required
> > to exclude VMD devices from getting the irq domain pointer set by
> > interrupt remapping.
> > 
> > Override the default bus token.
> > 
> > Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> > Acked-by: Bjorn Helgaas <bhelgaas@xxxxxxxxxx>
> >  drivers/pci/controller/vmd.c |    6 ++++++
> >  1 file changed, 6 insertions(+)
> > 
> > +++ b/drivers/pci/controller/vmd.c
> > @@ -579,6 +579,12 @@ static int vmd_enable_domain(struct vmd_
> >  		return -ENODEV;
> >  	}
> >  
> > +	/*
> > +	 * Override the irq domain bus token so the domain can be distinguished
> > +	 * from a regular PCI/MSI domain.
> > +	 */
> > +	irq_domain_update_bus_token(vmd->irq_domain, DOMAIN_BUS_VMD_MSI);
> > +
> 
> Having the non-transparent-bridge hold a MSI table and
> multiplex/de-multiplex IRQs looks like another good use case for
> something close to pci_subdevice_msi_create_irq_domain()?
> 
> If each device could have its own internal MSI-X table programmed
> properly things would work alot better. Disable capture/remap of the
> MSI range in the NTB.
We can disable the capture and remap in newer devices so we don't even
need the irq domain. Legacy VMD will automatically remap based on the
APIC dest bits in the MSI address.

I would however like to determine if the MSI data bits could be used
for individual devices to have the IRQ domain subsystem demultiplex the
virq from that and eliminate the IRQ list iteration.

A concern I have about that scheme is virtualization as I think those
bits are used to route to the virtual vector.

> 
> Then each device could have a proper non-multiplexed interrupt
> delivered to the CPU.. Affinity would work properly, no need to share
> IRQs (eg vmd_irq() goes away)/etc.
> 
> Something for the VMD maintainers to think about at least.
> 
> As I hear more about NTB these days a full MSI scheme for NTB seems
> interesting to have in the PCI-E core code..
> 
> Jason