From: Isaku Yamahata <isaku.yamahata@intel.com>

Add a flag, KVM_DEBUGREG_AUTO_SWITCHED_GUEST, to skip saving/restoring DRs
irrespective of any other flags.  TDX-SEAM unconditionally saves and
restores guest DRs and reset to architectural INIT state on TD exit.
So, KVM needs to save host DRs before TD enter without restoring guest DRs
and restore host DRs after TD exit.

Opportunistically convert the KVM_DEBUGREG_* definitions to use BIT().

Reported-by: Xiaoyao Li <xiaoyao.li@intel.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Co-developed-by: Chao Gao <chao.gao@intel.com>
Signed-off-by: Chao Gao <chao.gao@intel.com>
Signed-off-by: Isaku Yamahata <isaku.yamahata@intel.com>
---
 arch/x86/include/asm/kvm_host.h | 10 ++++++++--
 arch/x86/kvm/vmx/tdx.c          |  1 +
 arch/x86/kvm/x86.c              |  3 ++-
 3 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index b6396d11139e..489374a57b66 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -535,8 +535,14 @@ struct kvm_pmu {
 struct kvm_pmu_ops;
 
 enum {
-	KVM_DEBUGREG_BP_ENABLED = 1,
-	KVM_DEBUGREG_WONT_EXIT = 2,
+	KVM_DEBUGREG_BP_ENABLED		= BIT(0),
+	KVM_DEBUGREG_WONT_EXIT		= BIT(1),
+	KVM_DEBUGREG_RELOAD		= BIT(2),
+	/*
+	 * Guest debug registers are saved/restored by hardware on exit from
+	 * or enter guest. KVM needn't switch them.
+	 */
+	KVM_DEBUGREG_AUTO_SWITCH	= BIT(3),
 };
 
 struct kvm_mtrr_range {
diff --git a/arch/x86/kvm/vmx/tdx.c b/arch/x86/kvm/vmx/tdx.c
index 690298fb99c7..3a0e826fbe0c 100644
--- a/arch/x86/kvm/vmx/tdx.c
+++ b/arch/x86/kvm/vmx/tdx.c
@@ -485,6 +485,7 @@ int tdx_vcpu_create(struct kvm_vcpu *vcpu)
 
 	vcpu->arch.efer = EFER_SCE | EFER_LME | EFER_LMA | EFER_NX;
 
+	vcpu->arch.switch_db_regs = KVM_DEBUGREG_AUTO_SWITCH;
 	vcpu->arch.cr0_guest_owned_bits = -1ul;
 	vcpu->arch.cr4_guest_owned_bits = -1ul;
 
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 45e8a02e99bf..89d04cd64cd0 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -10084,7 +10084,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	if (vcpu->arch.guest_fpu.xfd_err)
 		wrmsrl(MSR_IA32_XFD_ERR, vcpu->arch.guest_fpu.xfd_err);
 
-	if (unlikely(vcpu->arch.switch_db_regs)) {
+	if (unlikely(vcpu->arch.switch_db_regs & ~KVM_DEBUGREG_AUTO_SWITCH)) {
 		set_debugreg(0, 7);
 		set_debugreg(vcpu->arch.eff_db[0], 0);
 		set_debugreg(vcpu->arch.eff_db[1], 1);
@@ -10126,6 +10126,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 	 */
 	if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT)) {
 		WARN_ON(vcpu->guest_debug & KVM_GUESTDBG_USE_HW_BP);
+		WARN_ON(vcpu->arch.switch_db_regs & KVM_DEBUGREG_AUTO_SWITCH);
 		static_call(kvm_x86_sync_dirty_debug_regs)(vcpu);
 		kvm_update_dr0123(vcpu);
 		kvm_update_dr7(vcpu);
-- 
2.25.1

On 3/4/22 20:49, isaku.yamahata@intel.com wrote:
>   	vcpu->arch.efer = EFER_SCE | EFER_LME | EFER_LMA | EFER_NX;
>   
> +	vcpu->arch.switch_db_regs = KVM_DEBUGREG_AUTO_SWITCH;
>   	vcpu->arch.cr0_guest_owned_bits = -1ul;
>   	vcpu->arch.cr4_guest_owned_bits = -1ul;
>   
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 45e8a02e99bf..89d04cd64cd0 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -10084,7 +10084,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>   	if (vcpu->arch.guest_fpu.xfd_err)
>   		wrmsrl(MSR_IA32_XFD_ERR, vcpu->arch.guest_fpu.xfd_err);
>   
> -	if (unlikely(vcpu->arch.switch_db_regs)) {
> +	if (unlikely(vcpu->arch.switch_db_regs & ~KVM_DEBUGREG_AUTO_SWITCH)) {
>   		set_debugreg(0, 7);
>   		set_debugreg(vcpu->arch.eff_db[0], 0);
>   		set_debugreg(vcpu->arch.eff_db[1], 1);

I'm confused.  I'm probably missing something obvious, but where does
KVM_DEBUGREG_AUTO_SWITCH affect the behavior of KVM?  vcpu_enter_guest
would still write %db0-%db3 if KVM_DEBUGREG_BP_ENABLED is set, for
example.

Do you mean:

	if (unlikely(vcpu->arch.switch_db_regs) &&
	    !unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_AUTO_SWITCH)) {

?

Paolo
  
> @@ -10126,6 +10126,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
>   	 */
>   	if (unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT)) {
>   		WARN_ON(vcpu->guest_debug & KVM_GUESTDBG_USE_HW_BP);
> +		WARN_ON(vcpu->arch.switch_db_regs & KVM_DEBUGREG_AUTO_SWITCH);
>   		static_call(kvm_x86_sync_dirty_debug_regs)(vcpu);
>   		kvm_update_dr0123(vcpu);
>   		kvm_update_dr7(vcpu);

On Tue, Apr 05, 2022, Paolo Bonzini wrote:
> On 3/4/22 20:49, isaku.yamahata@intel.com wrote:
> >   	vcpu->arch.efer = EFER_SCE | EFER_LME | EFER_LMA | EFER_NX;
> > +	vcpu->arch.switch_db_regs = KVM_DEBUGREG_AUTO_SWITCH;
> >   	vcpu->arch.cr0_guest_owned_bits = -1ul;
> >   	vcpu->arch.cr4_guest_owned_bits = -1ul;
> > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> > index 45e8a02e99bf..89d04cd64cd0 100644
> > --- a/arch/x86/kvm/x86.c
> > +++ b/arch/x86/kvm/x86.c
> > @@ -10084,7 +10084,7 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
> >   	if (vcpu->arch.guest_fpu.xfd_err)
> >   		wrmsrl(MSR_IA32_XFD_ERR, vcpu->arch.guest_fpu.xfd_err);
> > -	if (unlikely(vcpu->arch.switch_db_regs)) {
> > +	if (unlikely(vcpu->arch.switch_db_regs & ~KVM_DEBUGREG_AUTO_SWITCH)) {
> >   		set_debugreg(0, 7);
> >   		set_debugreg(vcpu->arch.eff_db[0], 0);
> >   		set_debugreg(vcpu->arch.eff_db[1], 1);
> 
> I'm confused.  I'm probably missing something obvious, but where does
> KVM_DEBUGREG_AUTO_SWITCH affect the behavior of KVM?  vcpu_enter_guest
> would still write %db0-%db3 if KVM_DEBUGREG_BP_ENABLED is set, for
> example.
> 
> Do you mean:
> 
> 	if (unlikely(vcpu->arch.switch_db_regs) &&
> 	    !unlikely(vcpu->arch.switch_db_regs & KVM_DEBUGREG_AUTO_SWITCH)) {
> 
> ?

No, you're confused there's a crucial chunk of this patch that's missing.  The
host restore path was originally

        if (hw_breakpoint_active() &&
            !(vcpu->arch.switch_db_regs & KVM_DEBUGREG_AUTO_SWITCHED))
                hw_breakpoint_restore();

The TDX module context switches the guest _and_ host debug registers.  It restores
the host DRs because it needs to write _something_ to hide guest state, so it might
as well restore the host values.  The above was an optmization to avoid rewriting
all debug registers.

However, this patch was written before commit f85d40160691 ("KVM: X86: Disable
hardware breakpoints unconditionally before kvm_x86->run()").  I suspect that when
rebasing to newer KVM, the Intel folks discovered that DR7 isn't getting restored
because the GUEST_DR7 in the SEAM VMCS will hold the zero'd value.

But rather than completely drop the flag, KVM can still avoid writing everything
except DR7, i.e.

	if (hw_breakpoint_active()) {
		if (!(vcpu->arch.switch_db_regs & KVM_DEBUGREG_AUTO_SWITCH))
			hw_breakpoint_restore();
		else
			set_debugreg(__this_cpu_read(cpu_dr7), 7);
	}

with a comment explaining that DR7 always needs to be restored because it's
preemptively cleared to play nice with the non-instrumentable sections.