Re: [kvm-unit-tests PATCH v2 0/4] Better backtraces for leaf functions

[Mathias Krause <minipli@grsecurity.net>]

On 18.11.25 05:04, Mathias Krause wrote:
> On 18.11.25 02:47, Mathias Krause wrote:
>> On 18.11.25 02:33, Mathias Krause wrote:
>>> On 17.11.25 23:19, Sean Christopherson wrote:
>>>> Spoke too soon :-(
>>>>
>>>> The x86 change breaks the realmode test.  I didn't try hard to debug, as that
>>>> test is brittle, e.g. see https://lore.kernel.org/all/20240604143507.1041901-1-pbonzini@redhat.com.
>>
>> Bleh, I just noticed, f01ea38a385a ("x86: Better backtraces for leaf
>> functions") broke vmx_sipi_signal_test too :(
>>
>> Looking into it!
> 
> Can't see it. The issue is, the call to hypercall(1) does return in
> x86/vmx.c:guest_entry() with f01ea38a385a applied but not w/o it. But
> that makes no sense to me whatsoever. :/

After some sleep it makes lot more sense. Well, not really, but it's a
race that f01ea38a385a, apparently, makes easier to hit.

> 
> Below debug diff shows the issue:
> 
> diff --git a/x86/vmx.c b/x86/vmx.c
> index c803eaa67ac6..e49668f5ff95 100644
> --- a/x86/vmx.c
> +++ b/x86/vmx.c
> @@ -806,4 +806,5 @@ asm(
>         "       mov $1, %edi\n\t"
>         "       call hypercall\n\t"
> +       "       ud2"
>  );
That hypercall call shouldn't return, as it's HYPERCALL_VMEXIT which
makes vmx_run() return and not re-enter the guest. But, apparently, it
does so and it does so because the guest/VMM shared variable
'hypercall_field' to communicate to the VMM what to do can also be
modified concurrently by the other CPU/guest #2 after it was written by
CPU/guest #1 but before read by the VMM. Bummer!

I'll send a fix in a few.

Thanks,
Mathias