Re: [PATCH kernel v3 2/3] KVM: SEV: Enable data breakpoints in SEV-ES

[Sean Christopherson <seanjc@google.com>]

On Fri, Jan 20, 2023, Alexey Kardashevskiy wrote:
> diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
> index 4826e6cc611b..61f2cad1cbaf 100644
> --- a/arch/x86/kvm/svm/svm.h
> +++ b/arch/x86/kvm/svm/svm.h
> @@ -389,6 +389,8 @@ static inline bool vmcb12_is_intercept(struct vmcb_ctrl_area_cached *control, u3
>  	return test_bit(bit, (unsigned long *)&control->intercepts);
>  }
>  
> +extern bool sev_es_is_debug_swap_enabled(void);
> +
>  static inline void set_dr_intercepts(struct vcpu_svm *svm)
>  {
>  	struct vmcb *vmcb = svm->vmcb01.ptr;
> @@ -410,8 +412,10 @@ static inline void set_dr_intercepts(struct vcpu_svm *svm)
>  		vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_WRITE);
>  	}
>  
> -	vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
> -	vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
> +	if (!sev_es_guest(svm->vcpu.kvm) || !sev_es_is_debug_swap_enabled()) {

Looking below, doesn't this do the wrong thing if set_dr_intercepts() is called
before SVM_SEV_FEAT_DEBUG_SWAP is set?  I.e. when this is called before LAUNCH_UPDATE?
Seems like this should check SVM_SEV_FEAT_DEBUG_SWAP in sev_features regardless
of when SVM_SEV_FEAT_DEBUG_SWAP is set.

And if KVM checks sev_features, then I _think_ we can avoid having to expose
sev_es_debug_swap_enabled to svm.{c,h} (though why on earth {set,clr}_dr_intercepts()
is in svm.h is another question for the future).

Follow-up question: does KVM _have_ to wait until KVM_SEV_LAUNCH_UPDATE_VMSA to
set the flag?

> +		vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
> +		vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
> +	}
>  
>  	recalc_intercepts(svm);
>  }
> @@ -422,8 +426,12 @@ static inline void clr_dr_intercepts(struct vcpu_svm *svm)
>  
>  	vmcb->control.intercepts[INTERCEPT_DR] = 0;
>  
> -	/* DR7 access must remain intercepted for an SEV-ES guest */
> -	if (sev_es_guest(svm->vcpu.kvm)) {
> +	/*
> +	 * DR7 access must remain intercepted for an SEV-ES guest unless DebugSwap
> +	 * (depends on NO_NESTED_DATA_BP) is enabled as otherwise a VM writing to DR7
> +	 * from the #DB handler may trigger infinite loop of #DB's.
> +	 */
> +	if (sev_es_guest(svm->vcpu.kvm) && !sev_es_is_debug_swap_enabled()) {
>  		vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
>  		vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
>  	}
>
> @@ -52,11 +53,21 @@ module_param_named(sev, sev_enabled, bool, 0444);
>  /* enable/disable SEV-ES support */
>  static bool sev_es_enabled = true;
>  module_param_named(sev_es, sev_es_enabled, bool, 0444);
> +
> +/* enable/disable SEV-ES DebugSwap support */
> +static bool sev_es_debug_swap_enabled = true;
> +module_param_named(debug_swap, sev_es_debug_swap_enabled, bool, 0644);

Module param needs 0444 permissions, i.e. shouldn't be writable after KVM is
loaded.  Though I don't know that providing a module param is warranted in this
case.  KVM provides module params for SEV and SEV-ES because there are legitimate
reasons to turn them off, but at a glance, I don't see why we'd want that for this
feature.

>  #else
>  #define sev_enabled false
>  #define sev_es_enabled false
> +#define sev_es_debug_swap false

This needs to be sev_es_debug_swap_enabled, otherwise things fall apart with
CONFIG_KVM_AMD_SEV=n.

arch/x86/kvm/svm/sev.c: In function ‘sev_es_is_debug_swap_enabled’:
arch/x86/kvm/svm/sev.c:69:16: error: ‘sev_es_debug_swap_enabled’ undeclared (first use in this function); did you mean ‘sev_es_is_debug_swap_enabled’?
   69 |         return sev_es_debug_swap_enabled;
      |                ^~~~~~~~~~~~~~~~~~~~~~~~~
      |                sev_es_is_debug_swap_enabled


>  #endif /* CONFIG_KVM_AMD_SEV */
>  
> +bool sev_es_is_debug_swap_enabled(void)
> +{
> +	return sev_es_debug_swap_enabled;
> +}

...

> @@ -604,6 +615,9 @@ static int sev_es_sync_vmsa(struct vcpu_svm *svm)
>  	save->xss  = svm->vcpu.arch.ia32_xss;
>  	save->dr6  = svm->vcpu.arch.dr6;
>  
> +	if (sev_es_is_debug_swap_enabled())
> +		save->sev_features |= SVM_SEV_FEAT_DEBUG_SWAP;
> +
>  	pr_debug("Virtual Machine Save Area (VMSA):\n");
>  	print_hex_dump_debug("", DUMP_PREFIX_NONE, 16, 1, save, sizeof(*save), false);
>  
> @@ -2249,6 +2263,9 @@ void __init sev_hardware_setup(void)
>  out:
>  	sev_enabled = sev_supported;
>  	sev_es_enabled = sev_es_supported;
> +	if (sev_es_debug_swap_enabled)
> +		sev_es_debug_swap_enabled = sev_es_enabled &&
> +			cpu_feature_enabled(X86_FEATURE_NO_NESTED_DATA_BP);

Slight preference for:

	if (!sev_es_enabled || !cpu_feature_enabled(X86_FEATURE_NO_NESTED_DATA_BP))
		sev_es_debug_swap_enabled = false;

KVM does short-circuit some checks on module param values, but usually only to
avoid additional setup.

>  #endif
>  }
>  
> @@ -3027,6 +3044,18 @@ void sev_es_prepare_switch_to_guest(struct sev_es_save_area *hostsa)
>  
>  	/* MSR_IA32_XSS is restored on VMEXIT, save the currnet host value */
>  	hostsa->xss = host_xss;
> +
> +	/* The DebugSwap SEV feature does Type B swaps of DR[0-3] */
> +	if (sev_es_is_debug_swap_enabled()) {
> +		hostsa->dr0 = native_get_debugreg(0);
> +		hostsa->dr1 = native_get_debugreg(1);
> +		hostsa->dr2 = native_get_debugreg(2);
> +		hostsa->dr3 = native_get_debugreg(3);
> +		hostsa->dr0_addr_mask = amd_get_dr_addr_mask(0);
> +		hostsa->dr1_addr_mask = amd_get_dr_addr_mask(1);
> +		hostsa->dr2_addr_mask = amd_get_dr_addr_mask(2);
> +		hostsa->dr3_addr_mask = amd_get_dr_addr_mask(3);
> +	}
>  }
>  
>  void sev_vcpu_deliver_sipi_vector(struct kvm_vcpu *vcpu, u8 vector)
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index 60c7c880266b..6c54a3c9d442 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -1190,7 +1190,8 @@ static void init_vmcb(struct kvm_vcpu *vcpu)
>  	set_exception_intercept(svm, UD_VECTOR);
>  	set_exception_intercept(svm, MC_VECTOR);
>  	set_exception_intercept(svm, AC_VECTOR);
> -	set_exception_intercept(svm, DB_VECTOR);
> +	if (!sev_es_is_debug_swap_enabled())
> +		set_exception_intercept(svm, DB_VECTOR);

This is wrong.  KVM needs to intercept #DBs when debugging non-SEV-ES VMs.
This _could_ be tied to X86_FEATURE_NO_NESTED_DATA_BP, but the KVM would need to
toggle the intercept depending on whether or not userspace wants to debug the
guest.

Similar to the DR7 interception, can this check sev_features directly?