On 27/11/18 17:27, Michael S. Tsirkin wrote: > On Tue, Nov 27, 2018 at 10:04:59AM +0100, Paolo Bonzini wrote: >> On 26/11/18 23:43, Michael S. Tsirkin wrote: >>> On Mon, Nov 26, 2018 at 02:20:38PM +0100, Paolo Bonzini wrote: >>>> vhost does not have to be supported only if KVM is present, in fact >>>> vhost-user does not even need any kind of kernel support. This series >>>> changes this. The rationale is that, when vhost-user-test.c will be >>>> converted to qgraph, it will be able to test vhost-user support >>>> for virtio-mmio backend even on x86. >>> >>> The reason for limiting it to KVM was very simple: >>> it has the same set of problems with ordering >>> as mttcg. >>> >>> So I guess it's fine but I think we must then limit it >>> to when tcg emits fence instructions. >>> >>> Otherwise we'll get subtle race conditions. >> >> Got it, I'll take a look. However, right now you _can_ use vhost with >> tcg as long as e.g. qemu-system-x86_64 is compiled with KVM support, so >> this patch is not changing anything in this respect; the bug is >> independent of this series. > > Well qemu built with KVM implies host architecture == guest > architecture, right? That should ensure race conditions do not happen in > a very hacky way.
It's not enough, event_idx requires smp_mb and single-threaded TCG does not emit them. It is enough for smp_rmb/smp_wmb however, that makes sense. Paolo
