Re: [kvm-unit-tests PATCH v3 04/20] x86: Dedup guest/host context switch of registers across SVM and VMX

[Sean Christopherson <seanjc@google.com>]

On Tue, May 19, 2026, Mathias Krause wrote:
> On 14.05.26 23:04, Sean Christopherson wrote:
> > Deduplicate the context switching of registers across VM-Enter<=>VM-Exit
> > between SVM and VMX.  The required functionality and implementations are
> > practically identical, literally the only difference is that SVM doesn't
> > need (or want) to manually swap RAX as SVM automatically swaps RAX at
> > VMRUN and #VMEXIT.
> > 
> > Opportunistically rename the structure to "guest_regs" to clarify its
> > purpose, and to avoid conflicts, e.g. with realmode's "struct regs".
> > 
> > Signed-off-by: Sean Christopherson <seanjc@google.com>
> > ---
> >  lib/x86/virt.h | 52 ++++++++++++++++++++++++++++++++++++++++++++++++++
> >  x86/svm.c      |  4 ++--
> >  x86/svm.h      | 47 ++++-----------------------------------------
> >  x86/vmx.c      |  8 ++++----
> >  x86/vmx.h      | 42 +---------------------------------------
> >  5 files changed, 63 insertions(+), 90 deletions(-)
> >  create mode 100644 lib/x86/virt.h
> > 
> > diff --git a/lib/x86/virt.h b/lib/x86/virt.h
> > new file mode 100644
> > index 00000000..ccc90c25
> > --- /dev/null
> > +++ b/lib/x86/virt.h
> > @@ -0,0 +1,52 @@
> > +#ifndef _x86_VIRT_H_
> > +#define _x86_VIRT_H_
> > +
> > +#include "libcflat.h"
> > +
> > +struct guest_regs {
> > +	u64 rax;
> > +	u64 rcx;
> > +	u64 rdx;
> > +	u64 rbx;
> > +	/*
> > +	 * Use RSP's index to hold CR3, as RSP isn't manually context switched
>                                    ^^^ that should be CR2...

This one should be CR2.  CR2 isn't context switch via the VMCS.  Ugh, but looking
at the usage, I think this has been dead code since it was introduced in commit
9d7eaa29 ("kvm-unit-tests : Basic architecture of VMX nested test case").  I was
just reverse engineering the intent when writing the comment, and didn't actually
check to see if KUT does what someone intended it to do.

nVMX likely works because no test takes page faults in both L1 and L2, i.e. either
L2 doesn't care about CR2, or L1 never clobbers it.

Hrm.  I'm leaning towards deleting my comment and maintaining the status quo, for
now.  As tempting as it is to fix the bug, nSVM doesn't need the help, there's
zero reason to context CR2 in assembly code, and I doubt a test will ever rely on
CR2 being context switched.  

> > +	 * by software in any relevant flows.
> > +	 */
> > +	u64 cr2;
>             ^^^ or this one cr3?
> 
> > +	u64 rbp;
> > +	u64 rsi;
> > +	u64 rdi;
> > +	u64 r8;
> > +	u64 r9;
> > +	u64 r10;
> > +	u64 r11;
> > +	u64 r12;
> > +	u64 r13;
> > +	u64 r14;
> > +	u64 r15;
> > +	u64 rflags;