On 9 April 2015 at 12:40, Paolo Bonzini <pbonz...@redhat.com> wrote: > > > On 09/04/2015 12:43, Peter Maydell wrote: >> > At this point, some memory barriers, basically. >> >> So what distinguishes a device that needs the memory barriers >> and does its accesses via dma_* from a device that doesn't and >> uses address_space_* or ld/st*_phys ? (Or for that matter a >> non-device that does memory accesses...) > > I don't know exactly, I didn't follow the discussion very much back > then. The memory barriers were fixing PPC bugs; PCI devices definitely > need them.
I suspect that actually we need barriers in a lot more places and there shouldn't really be a separate set of APIs here; it's just that nobody's tried using the wider set of devices in PPC KVM systems. -- PMM