The first step in making nested SVM production-ready is to make sure that all features are implemented and well-tested. To make this tractable, we will initially be limiting the “supported” range of nested virt to a specific subset of host and guest features. This document describes the criteria for deciding on features, and the rationale behind each feature.
For AMD, all virtualization-related features can be found in CPUID leaf 8000000A:edx
Processor support: At a minimum we want to support processors from the last 5 years. All things being equal, we’d prefer to cover older processors than not. Bits 0:7 were available in the very earliest processors; and even through bit 15 we should be pretty good support-wise.
Faithfulness to hardware: We need the behavior of the “virtual cpu” from the L1 hypervisor’s perspective to be as close as possible to the original hardware. In particular, the behavior of the hardware on error paths 1) is not easy to understand or test, 2) can be the source of surprising vulnerabiliies. (See XSA-7 for an example of a case where subtle error-handling differences can open up a privilege escalation.) We should avoid emulating any bit of the hardware with complex error paths if we can at all help it.
Cost of implementation: We want to minimize the cost of implementation (where this includes bringing an existing sub-par implementation up to speed). All things being equal, we’ll favor a configuration which does not require any new implementation.
Performance: All things being equal, we’d prefer to choose a set of L0 / L1 CPUID bits that are faster than slower.
0 NP Nested Paging: Required both for L0 and L1.
Based primarily on faithfulness and performance, as well as potential cost of implementation. Available on earliest hardware, so no compatibility issues.
1 LbrVirt LBR / debugging virtualization: Require for L0 and L1.
For L0 this is required for performance: There’s no way to tell the guests not to use the LBR-related registers; and if the guest does, then you have to save and restore all LBR-related registers on context switch, which is prohibitive. Furthermore, the additional emulation risks a security-relevant difference to come up.
Providing it to L1 when we have it in L0 is basically free, and already implemented.
Just require it and provide it.
2 SVML SVM Lock: Not required for L0, not provided to L1
Seems to be aboult enabling an operating system to prevent “blue pill” attacks against itself.
Xen doesn’t use it, nor provide it; so it would need to be implementend. The best way to protect a guest OS is to leave nested virt disabled in the tools.
3 NRIPS NRIP Save: Require for both L0 and L1
If NRIPS is not present, the software interrupt injection functionality can’t be used; and Xen has to emulate it. That’s another source of potential security issues. If hardware supports it, then providing it to guest is basically free.
4 TscRateMsr: Not required by L0, not provided to L1
The main putative use for this would be trying to maintain an invariant TSC across cores with different clock speeds, or after a migrate. Unlike others, this doesn’t have an error path to worry about compatibility-wise; and according to tests done when nestedSVM was first implemented, it’s actually faster to emliate TscRateMSR in the L0 hypervisor than for L1 to attempt to emulate it itself.
However, using this properly in L0 will take some implementation effort; and composing it properly with L1 will take even more effort. Just leave it off for now.
5 VmcbClean: VMCB Clean Bits: Not required by L0, provide to L1
This is a pure optimization, both on the side of the L0 and L1. The implementaiton for L1 is entirely Xen-side, so can be provided even on hardware that doesn’t provide it. And it’s purely an optimization, so could be “implemented” by ignoring the bits entirely.
As such, we don’t need to require it for L0; and as it’s already implemented, no reason not to provide it to L1. Before this feature was available those bits were marked SBZ (“should be zero”); setting them was already advertised to cause unpredictable behavior.
6 FlushByAsid: Require for L0, provide to L1
This is cheap and easy to use for L0 and to provide to the L1; there’s no reson not to just pass it through.
7 DecodeAssists: Require for L0, provide to L1
Using it in L0 reduces the chance that we’ll make some sort of error in the decode path. And if hardware supports it, it’s easy enough to provide to the L1.