On 5/10/2022 11:51 PM, Paolo Bonzini wrote:
On 5/6/22 05:33, Yang Weijiang wrote:
Per spec:"IA32_LBR_CTL.LBREn is saved and cleared on #SMI, and restored
on RSM. On a warm reset, all LBR MSRs, including IA32_LBR_DEPTH, have their
values preserved. However, IA32_LBR_CTL.LBREn is cleared to 0, disabling
LBRs." So clear Arch LBREn bit on #SMI and restore it on RSM manully, also
clear the bit when guest does warm reset.
Signed-off-by: Yang Weijiang <weijiang.yang@xxxxxxxxx>
---
arch/x86/kvm/vmx/vmx.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index 6d6ee9cf82f5..b38f58868905 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -4593,6 +4593,8 @@ static void vmx_vcpu_reset(struct kvm_vcpu *vcpu, bool init_event)
if (!init_event) {
if (static_cpu_has(X86_FEATURE_ARCH_LBR))
vmcs_write64(GUEST_IA32_LBR_CTL, 0);
+ } else {
+ flip_arch_lbr_ctl(vcpu, false);
}
}
@@ -7704,6 +7706,7 @@ static int vmx_enter_smm(struct kvm_vcpu *vcpu, char *smstate)
vmx->nested.smm.vmxon = vmx->nested.vmxon;
vmx->nested.vmxon = false;
vmx_clear_hlt(vcpu);
+ flip_arch_lbr_ctl(vcpu, false);
return 0;
}
@@ -7725,6 +7728,7 @@ static int vmx_leave_smm(struct kvm_vcpu *vcpu, const char *smstate)
vmx->nested.nested_run_pending = 1;
vmx->nested.smm.guest_mode = false;
}
+ flip_arch_lbr_ctl(vcpu, true);
return 0;
}
This is incorrect, you hare not saving/restoring the actual value of
LBREn (which is "lbr_desc->event != NULL"). Therefore, a migration
while in SMM would lose the value of LBREn = true.
Instead of using flip_arch_lbr_ctl, SMM should save the value of the MSR
in kvm_x86_ops->enter_smm, and restore it in kvm_x86_ops->leave_smm
(feel free to do it only if guest_cpuid_has(vcpu, X86_FEATURE_LM), i.e.
the 32-bit case can be ignored).
In the case of migration in SMM, I assume kvm_x86_ops->enter_smm()
called in source side
and kvm_x86_ops->leave_smm() is called at destination, then should the
saved LBREn be transferred
to destination too? The destination can rely on the bit to defer setting
LBREn bit in
VMCS until kvm_x86_ops->leave_smm() is called. is it good? thanks!
Paolo
