Re: [PATCH 5/5] KVM: x86: Defer runtime updates of dynamic CPUID bits until CPUID emulation

[Sean Christopherson <seanjc@google.com>]

On Fri, Nov 28, 2025, Igor Mammedov wrote:
> On Tue, 10 Dec 2024 17:33:02 -0800
> Sean Christopherson <seanjc@google.com> wrote:
> 
> Sean,
> 
> this patch broke vCPU hotplug (still broken with current master),
> after repeated plug/unplug of the same vCPU in a loop, QEMU exits
> due to error in vcpu initialization:
> 
>     r = kvm_vcpu_ioctl(cs, KVM_SET_CPUID2, &cpuid_data);                         
>     if (r) {                                                                     
>         goto fail;                                                               
>     }
> 
> Reproducer (host in question is Haswell but it's been seen on other hosts as well):
> for it to trigger the issue it must be Q35 machine with UEFI firmware
> (the rest doesn't seem to matter)

Gah, sorry.  I managed to handle KVM_GET_CPUID2, so I suspect I thought the update
in kvm_cpuid_check_equal() would take care of things, but that only operates on
the new entries.

Can you test the below?  In the meantime, I'll see if I can enhance the CPUID
selftest to detect the issue and verify the fix.

diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
index d563a948318b..dd6534419074 100644
--- a/arch/x86/kvm/cpuid.c
+++ b/arch/x86/kvm/cpuid.c
@@ -509,11 +509,18 @@ static int kvm_set_cpuid(struct kvm_vcpu *vcpu, struct kvm_cpuid_entry2 *e2,
        u32 vcpu_caps[NR_KVM_CPU_CAPS];
        int r;
 
+       /*
+        * Apply pending runtime CPUID updates to the current CPUID entries to
+        * avoid false positives due mismatches on KVM-owned feature flags.
+        */
+       if (vcpu->arch.cpuid_dynamic_bits_dirty)
+               kvm_update_cpuid_runtime(vcpu);
+
        /*
         * Swap the existing (old) entries with the incoming (new) entries in
         * order to massage the new entries, e.g. to account for dynamic bits
-        * that KVM controls, without clobbering the current guest CPUID, which
-        * KVM needs to preserve in order to unwind on failure.
+        * that KVM controls, without losing the current guest CPUID, which KVM
+        * needs to preserve in order to unwind on failure.
         *
         * Similarly, save the vCPU's current cpu_caps so that the capabilities
         * can be updated alongside the CPUID entries when performing runtime