2016-02-19 12:43+0100, Jan Kiszka: > On 2016-02-19 12:34, Peter Xu wrote: >> On Fri, Feb 19, 2016 at 10:58:12AM +0100, Paolo Bonzini wrote: >>> >>> The source ID can be passed to the IOMMU using the MemTxAttrs mechanism. >> >> I see that MemTxAttrs is not enabled yet? It's dropped in all the >> accessors like memory_region_write_accessor()? > > That would be a pity. How much work remains to fix that? Putting Peter > Maydell on CC who apparently started this extension.
memory_region_dispatch_write() uses write_with_attrs, so I think we only need to use one for MSI region and add attributes to direct writers (like HPET and IOAPIC, because normal PCI devices should be passing correct attributes). > ... >>> FWIW, Radim was thinking of interrupt remapping in the kvm-ioapic, which >>> we have decided to set aside. > >> Does it mean that we are planning not to support kernel_irqchip for >> IR? Any quick reason? > > The reason to split up the kernel irqchip is to reduce code complexity > and, thus, attack surface of the kernel-side interface towards the > guest. So, extending the "classic" in-kernel support with yet another > feature is not really supporting that goal. A fear of regressions after forcing people to switch from kernel ioapic to qemu ioapic is the only argument for in-kernel support. We'd like to use split irqchip even without IR.