On Tue, Oct 18, 2011 at 12:46:50PM +1100, David Gibson wrote: > On Sun, Oct 16, 2011 at 03:15:53PM +0200, Avi Kivity wrote: > > On 10/14/2011 04:14 AM, David Gibson wrote: > > > > Virtio is a very, very special case. virtio requires coherent RAM > > > > access. > > > > > > Right. Virtio's access to memory is *not* emulated PCI DMA, it's > > > god-like hypervisor access to guest system memory. It should > > > correctly bypass any IOMMU, and so should remain as > > > cpu_physical_memory_rw() or the atomic accessors, rather than being > > > converted to this new API. > > > > virtio should definitely not bypass an iommu. > > So, I just had a chat with Rusty about this. Perhaps it shouldn't, > but it does. The spec is in terms of guest physical addresses, not > bus/DMA addresses, and more to the point the Linux driver does *not* > do the necessary dma_map() and unmap operations to treat this as a PCI > DMA. So like it or not, god-like hypervisor access rather than > emulated PCI DMA is what it does.
Fine, but I'm convinced virtio is not unique in that it wants atomic accesses. I just looked at hw/rtl8139.c as one example. It uses a high bit in a 32 bit register to signal descriptor ownership. Thus we need to read that bit first, or read the register atomically. Current code does cpu_physical_memory_read which does neither of these things, but it seems to be a bug. An atomic load would be the best solution. -- MST
