Guest hangs in specific configurations (KVM_X86_DISABLE_EXITS_HLT) are 
reported and the issue was bisected to commit ee3a5f9e3d9b ("KVM: x86: Do
runtime CPUID update before updating vcpu->arch.cpuid_entries") which, of
course, carries "No functional change intended" blurb. Turns out, moving
__kvm_update_cpuid_runtime() earlier in kvm_set_cpuid() to tweak the 
incoming CPUID data before checking it wasn't innocent as 
KVM_FEATURE_PV_UNHALT reset logic relies on cached KVM CPUID base which
gets updated later.

I was not able to reproduce the issue with QEMU myself so I wrote a
selftest to show the problem.

Vitaly Kuznetsov (3):
  KVM: x86: Introduce __kvm_get_hypervisor_cpuid() helper
  KVM: x86: Use actual kvm_cpuid.base for clearing KVM_FEATURE_PV_UNHALT
  KVM: selftests: Check that KVM_FEATURE_PV_UNHALT is cleared with
    KVM_X86_DISABLE_EXITS_HLT

 arch/x86/kvm/cpuid.c                          | 42 ++++++++++++-------
 .../selftests/kvm/x86_64/kvm_pv_test.c        | 42 +++++++++++++++++++
 2 files changed, 68 insertions(+), 16 deletions(-)


base-commit: 0cbca1bf44a0b8666c91ce3438f235c6fe70fbf1
-- 
2.43.0

On Wed, 28 Feb 2024 11:18:34 +0100, Vitaly Kuznetsov wrote:
> Guest hangs in specific configurations (KVM_X86_DISABLE_EXITS_HLT) are
> reported and the issue was bisected to commit ee3a5f9e3d9b ("KVM: x86: Do
> runtime CPUID update before updating vcpu->arch.cpuid_entries") which, of
> course, carries "No functional change intended" blurb. Turns out, moving
> __kvm_update_cpuid_runtime() earlier in kvm_set_cpuid() to tweak the
> incoming CPUID data before checking it wasn't innocent as
> KVM_FEATURE_PV_UNHALT reset logic relies on cached KVM CPUID base which
> gets updated later.
> 
> [...]

Applied to kvm-x86 hyperv.  I won't send a pull request for this until next week,
but I do plan on landing it in 6.9.  Holler if the selftests tweaks look wrong
(or you just don't like them).  Thanks!

[1/3] KVM: x86: Introduce __kvm_get_hypervisor_cpuid() helper
      https://github.com/kvm-x86/linux/commit/92e82cf632e8
[2/3] KVM: x86: Use actual kvm_cpuid.base for clearing KVM_FEATURE_PV_UNHALT
      https://github.com/kvm-x86/linux/commit/4736d85f0d18
[3/3] KVM: selftests: Check that PV_UNHALT is cleared when HLT exiting is disabled
      https://github.com/kvm-x86/linux/commit/c2585047c8e1

--
https://github.com/kvm-x86/linux/tree/next

Sean Christopherson <seanjc@google.com> writes:

> On Wed, 28 Feb 2024 11:18:34 +0100, Vitaly Kuznetsov wrote:
>> Guest hangs in specific configurations (KVM_X86_DISABLE_EXITS_HLT) are
>> reported and the issue was bisected to commit ee3a5f9e3d9b ("KVM: x86: Do
>> runtime CPUID update before updating vcpu->arch.cpuid_entries") which, of
>> course, carries "No functional change intended" blurb. Turns out, moving
>> __kvm_update_cpuid_runtime() earlier in kvm_set_cpuid() to tweak the
>> incoming CPUID data before checking it wasn't innocent as
>> KVM_FEATURE_PV_UNHALT reset logic relies on cached KVM CPUID base which
>> gets updated later.
>> 
>> [...]
>
> Applied to kvm-x86 hyperv.  I won't send a pull request for this until next week,
> but I do plan on landing it in 6.9.  Holler if the selftests tweaks look wrong
> (or you just don't like them).

Looks great, thanks :-)

I was also considering introducing 'vcpu_cpuid_has()' first but then I
succumbed to my laziness and decided to postpone it until we have a
*second* user in the tree).

-- 
Vitaly