> On 8 Nov 2018, at 11:50, Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote: > > On 08/11/2018 01:45, Jim Mattson wrote: >> I have no attachments to the current design. I had used a data[] blob, >> because I didn't think userspace would have any need to know what was >> in there. However, I am now seeing the error of my ways. For example, >> the userspace instruction emulator needs to know the contents of the >> vmcs12 to emulate instructions when in guest mode. > > Yeah, we're probably going to have to document the KVM vmcs12 structure, > possibly moving it to uapi. But that's a different thing from > save/restore state, which can use the 4K or 8K data[] blob. > > Paolo But regardless of if we document vmcs12 or not, the current blob we have today should be separated to well-defined blobs/structs (cached_vmcs12 and cached_shadow_vmcs12) and each blob should have a relevant flag that specifies it is valid (saved by kernel or requested to be restored by userspace). Additional future nested-state should be added as additional well-defined blobs/structs with appropriate flags. Then, in QEMU, each such well-defined blob/struct should have it’s own subsection with a relevant .needed() method. This will allow us to preserve required backwards compatibility. Agreed? -Liran
