Re: [PATCH v4 11/12] KVM: selftests: Support running stress save+restore and #PF test in L2

[Sean Christopherson <seanjc@google.com>]

On Mon, Jul 27, 2026, Yosry Ahmed wrote:
> Extend the stress test to allow running the access+#PF code in L2
> instead of L1 by adding proper L1 guest code to bootstrap L2. By
> default, the test runs in L2 after running in L1 if nested is supported.
> 
> Assisted-by: Gemini:gemini-3.1-pro
> Signed-off-by: Yosry Ahmed <yosry@kernel.org>
> ---
>  .../kvm/x86/save_restore_pf_stress_test.c     | 61 ++++++++++++++++++-
>  1 file changed, 58 insertions(+), 3 deletions(-)
> 
> diff --git a/tools/testing/selftests/kvm/x86/save_restore_pf_stress_test.c b/tools/testing/selftests/kvm/x86/save_restore_pf_stress_test.c
> index 664ed280b2e76..0e5ddeb5af444 100644
> --- a/tools/testing/selftests/kvm/x86/save_restore_pf_stress_test.c
> +++ b/tools/testing/selftests/kvm/x86/save_restore_pf_stress_test.c
> @@ -8,10 +8,13 @@
>  #include <pthread.h>
>  #include <signal.h>
>  #include <unistd.h>
> +#include <getopt.h>
>  
>  #include "test_util.h"
>  #include "kvm_util.h"
>  #include "processor.h"
> +#include "svm_util.h"
> +#include "vmx.h"
>  
>  #define NR_ITERATIONS		500
>  
> @@ -81,6 +84,36 @@ static void guest_access_memory(void *arg)
>  	}
>  }
>  
> +static void l1_svm_code(struct svm_test_data *svm)
> +{
> +	generic_svm_setup(svm, guest_access_memory);
> +	run_guest(svm->vmcb, svm->vmcb_gpa);
> +	GUEST_ASSERT(false);
> +}
> +
> +static void l1_vmx_code(struct vmx_pages *vmx)
> +{
> +	GUEST_ASSERT(prepare_for_vmx_operation(vmx));
> +	GUEST_ASSERT(load_vmcs(vmx));
> +	prepare_vmcs(vmx, guest_access_memory);
> +
> +	/* Ignore any #PF */

This comment is rather ambiguous, especially when the #UD interception comes
along, because unless the reader is reading carefully and sees the PFEC_MASK
and PFEC_MATCH logic below, it's quite easy to read this as "intercept #PF to
ignore them".

I don't see any reason to write the code this way.  Yeah yeah, it's the SDM's
suggested way to ignore #PFs, but IMO that's unnecessarily convoluted.  The more
obvious way is to leave all fields '0'.

In other words, just delete this entire block of code and rely on the default
VMX allocation logic to disable interception of all exceptions.

Oh, damnit.  Argh.  I see why you need this.  init_vmcs_control_fields() is buggy
and actually configures #PF for interception.  Apparently Paolo didn't read the
"If there is inequality, the meaning of that bit is reversed (for example, a VM
exit occurs if that bit is clear)." part :-D

So, let's fix that in a prep patch (AFAICT, no existing tests cares about #PF
interception) so that this test doesn't have to carry confusing code, and because
intercpeting #PF but nothing else by default is bound to cause problems.

diff --git tools/testing/selftests/kvm/lib/x86/vmx.c tools/testing/selftests/kvm/lib/x86/vmx.c
index cd09c9de4485..089e1a8af53f 100644
--- tools/testing/selftests/kvm/lib/x86/vmx.c
+++ tools/testing/selftests/kvm/lib/x86/vmx.c
@@ -232,7 +232,7 @@ static inline void init_vmcs_control_fields(struct vmx_pages *vmx)
 
        vmwrite(EXCEPTION_BITMAP, 0);
        vmwrite(PAGE_FAULT_ERROR_CODE_MASK, 0);
-       vmwrite(PAGE_FAULT_ERROR_CODE_MATCH, -1); /* Never match */
+       vmwrite(PAGE_FAULT_ERROR_CODE_MATCH, 0);
        vmwrite(CR3_TARGET_COUNT, 0);
        vmwrite(VM_EXIT_CONTROLS, rdmsr(MSR_IA32_VMX_EXIT_CTLS) |
                VM_EXIT_HOST_ADDR_SPACE_SIZE);    /* 64-bit host */

> +	GUEST_ASSERT(!vmwrite(EXCEPTION_BITMAP, BIT(PF_VECTOR)));
> +	GUEST_ASSERT(!vmwrite(PAGE_FAULT_ERROR_CODE_MASK, 0));
> +	GUEST_ASSERT(!vmwrite(PAGE_FAULT_ERROR_CODE_MATCH, -1));
> +
> +	GUEST_ASSERT(!vmlaunch());
> +	GUEST_ASSERT(false);
> +}