On 09/06/2011 11:28 AM, Avi Kivity wrote:
Actually Michael is right. The implementation is correct on x86, though wrong anywhere else (perhaps s390?). On those architectures you do not need rmb() and wmb().Are we sure? Nothing prevents the guest from using weakly-ordered writes, is there? For example, using MOVNTDQ. Although in that case the guest is probably required to issue an SFENCE.
Yes, that's the guest problem. You cannot do an SFENCE on the guest's behalf anyway.
Paolo
