On 05/10/2022 23:02, Suraj Jitindar Singh wrote:
tl;dr: The existing mitigation for eIBRS PBRSB predictions uses an INT3 to ensure a call instruction retires before a following unbalanced RET.
No it doesn't. The INT3 is transient. The existing mitigation uses an LFENCE.
Replace this with a WRMSR serialising instruction which has a lower performance penalty.
What is "this"? The INT3^W LFENCE?
WRMSR is not lower overhead than an LFENCE.
== Solution ==
The WRMSR instruction can be used as a speculation barrier and a serialising instruction. Use this on the VM exit path instead to ensure that a CALL instruction (in this case the call to vmx_spec_ctrl_restore_host) has retired before the prediction of a following unbalanced RET.
While both of these sentences are true statements, you've missed the necessary safety property.
One CALL has to retire before *any* RET can execute.
There are several ways the frontend can end up eventually consuming the bad RSB entry; they all stem from an execute (not prediction) of the next RET instruction.
As to the change, ...
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c index c9b49a09e6b5..fdcd8e10c2ab 100644 --- a/arch/x86/kvm/vmx/vmx.c +++ b/arch/x86/kvm/vmx/vmx.c @@ -7049,8 +7049,13 @@ void noinstr vmx_spec_ctrl_restore_host(struct vcpu_vmx *vmx,
... out of context above this hunk is:
if (!cpu_feature_enabled(X86_FEATURE_MSR_SPEC_CTRL)) return;
meaning that there is a return instruction which is programmatically reachable ahead of the WRMSR.
Whether it is speculatively reachable depends on whether the frontend can see through the _static_cpu_has(), as well as X86_FEATURE_MSR_SPEC_CTRL never becoming compile time evaluable.
There is also a second latent bug, to do with the code generation for this_cpu_read(x86_spec_ctrl_current).
It's worth saying that, in principle, this optimisation is safe, but pretty much all the discussion about it is wrong. Here is one I prepared earlier: https://lore.kernel.org/xen-devel/20220809170016.25148-3-andrew.cooper3@citr...
OTOH, below the hunk in question, there's a barrier_nospec() which is giving you the actual projection you need (subject to latent and/or code layout bugs), irrespective of the extra WRMSR.
~Andrew