On 06/18/2018 08:13 AM, Keno Fischer wrote: >>> 4) Catch the fault thrown by xsaves/xrestors in this situation, update >>> XCR0, redo the xsaves/restores, put XCR0 back and continue >>> execution after the faulting instruction. >> >> I'm worried about the kernel pieces that go digging in the XSAVE data >> getting confused more than the hardware getting confused. > > So you prefer this option? If so, I can try to have a go at implementing it > this way and seeing if I run into any trouble.
No, I'm saying that depending on faults is not a viable solution. We are not guaranteed to get faults in all the cases you would need to fix up. XSAVE*/XRSTOR* are not even *called* in some of those cases. >>> At least currently, it is my understanding that `xfeatures_mask` only has >>> user features, am I mistaken about that? >> >> We're slowing adding supervisor support. I think accounting for >> supervisor features is a requirement for any new XSAVE code. > > Sure, I don't think this is in any way incompatible with that (though > probably also informs that we want to keep the memory layout the > same if possible). I think you've tried to simplify your implementation by ignoring features, like holes. However, the existing implementation actually *does* handle those things and we've spent a significant amount of time ensuring that it works, despite the fact that you can't buy an off-the-shelf CPU that creates a hole without help from a hypervisor today.
