On 3/23/23 12:40, Sean Christopherson wrote:
On Fri, Feb 03, 2023, Alexey Kardashevskiy wrote:
While at this, move set_/clr_dr_intercepts to .c and move #DB intercept
next to DR7 intercept.
Please do non-trivial code movement in separate patches unless the functional change
is trivial. Moving and changing at the same time makes the patch difficult to review.
@@ -52,9 +53,14 @@ 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);
Needs to be 0444, otherwise userspace can turn on the knob after KVM is loaded,
which would allow enabling the feature on unsupported platforms, amongst many
other problems.
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..f8e222bee22a 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -671,6 +671,65 @@ static int svm_cpu_init(int cpu)
}
+static void set_dr_intercepts(struct vcpu_svm *svm)
+{
+ struct vmcb *vmcb = svm->vmcb01.ptr;
+ bool intercept;
+
+ if (!sev_es_guest(svm->vcpu.kvm)) {
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_WRITE);
+ }
+
+ if (sev_es_guest(svm->vcpu.kvm)) {
+ struct sev_es_save_area *save = svm->sev_es.vmsa;
+
+ intercept = !(save->sev_features & SVM_SEV_FEAT_DEBUG_SWAP);
Blech, the VMCB vs. SEV and SEV-ES code is a mess. E.g. init_vmcb() does
/*
* Guest access to VMware backdoor ports could legitimately
* trigger #GP because of TSS I/O permission bitmap.
* We intercept those #GP and allow access to them anyway
* as VMware does. Don't intercept #GP for SEV guests as KVM can't
* decrypt guest memory to decode the faulting instruction.
*/
if (enable_vmware_backdoor && !sev_guest(vcpu->kvm))
set_exception_intercept(svm, GP_VECTOR);
but then sev_es_init_vmcb() also does:
/* No support for enable_vmware_backdoor */
clr_exception_intercept(svm, GP_VECTOR);
DR interception is a similar trainwreck. svm_sync_dirty_debug_regs() bails if
guest_state_protected is true, i.e. is a nop for SEV-ES guests, but only after
the vCPU has done LAUNCH_UPDATE_VMSA. IIUC, that's nonsensical because even before
guest state is encrypted, #DB will be reflected as #VC into the guest. And, again
A guest can't run before the LAUNCH_UPDATE process is complete, so there
can't be a #VC before guest_state_proteced is true.
IIUC, except for DR7, DRs are never intercepted for SEV-ES guests and so trying
to debug from the host is futile as the guest can clobber DRs at any time.
Similarly, flowing into dr_interception() on an SEV-ES VMGEXITis just dumb. KVM
_knows_ it can't give the guest control of DR7, but it mucks with the intercepts
anyways. That the GHCB spec even allows SVM_EXIT_{READ,WRITE}_DR7 is just asinine,
but that's a moot point. Anyways, the GHCB spec's "suggestion" effectively says
KVM's responsibility is purely to make a read of DR7 return the last written value.
That's not KVM's responsibility, that is the responsibility of the guest
#VC handler. So a DR7 read, while intercepted, should never get to KVM.
And of course KVM's disaster of a flow doesn't even do that unless the host is
debugging the guest.
Currently, hardware debug traps aren’t supported for an SEV-ES guest. The hypervisor
must set the intercept for both read and write of the debug control register (DR7).
With the intercepts in place, the #VC handler will be invoked when the guest accesses
DR7. For a write to DR7, the #VC handler should perform Standard VMGExit processing.
The #VC handler must not update the actual DR7 register, but rather it should cache
the DR7 value being written.
I bring this up because of the subtle dependency that checking SVM_SEV_FEAT_DEBUG_SWAP
creates: set_dr_intercepts() needs to be called after sev_init_vmcb(). I believe
this approach also fails to handle intrahost migration; at the very least, what
exactly will happen when sev_migrate_from() invokes sev_init_vmcb() is unclear.
And I really don't want to pile even more gunk on top of the existing mess.
So, can you (and by "you" I really mean "the folks at AMD working on SEV stuff")
start with the below diff (not intended to be a single patch), disallow
kvm_arch_vcpu_ioctl_set_guest_debug() entirely for SEV-ES guests (will likely
take some back and forth to figure out how we want to do this), and then fill
in the blanks? I.e. get KVM to a state where all the intercept shenanigans for
SEV and SEV-ES are reasonably contained in sev.c, and then enable the debug_swap
stuff on top?
diff --git a/arch/x86/kvm/svm/sev.c b/arch/x86/kvm/svm/sev.c
index c25aeb550cd9..ff7a4d68731c 100644
--- a/arch/x86/kvm/svm/sev.c
+++ b/arch/x86/kvm/svm/sev.c
@@ -2968,8 +2968,7 @@ static void sev_es_init_vmcb(struct vcpu_svm *svm)
svm_set_intercept(svm, TRAP_CR4_WRITE);
svm_set_intercept(svm, TRAP_CR8_WRITE);
- /* No support for enable_vmware_backdoor */
- clr_exception_intercept(svm, GP_VECTOR);
+ <debug register stuff goes here>
/* Can't intercept XSETBV, HV can't modify XCR0 directly */
svm_clr_intercept(svm, INTERCEPT_XSETBV);
@@ -2996,6 +2995,12 @@ void sev_init_vmcb(struct vcpu_svm *svm)
svm->vmcb->control.nested_ctl |= SVM_NESTED_CTL_SEV_ENABLE;
clr_exception_intercept(svm, UD_VECTOR);
+ /*
+ * Don't intercept #GP for SEV guests, e.g. for the VMware backdoor, as
+ * KVM can't decrypt guest memory to decode the faulting instruction.
+ */
+ clr_exception_intercept(svm, GP_VECTOR);
+
if (sev_es_guest(svm->vcpu.kvm))
sev_es_init_vmcb(svm);
}
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index e0ec95f1f068..89753d7fd821 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -1209,10 +1209,9 @@ static void init_vmcb(struct kvm_vcpu *vcpu)
* Guest access to VMware backdoor ports could legitimately
* trigger #GP because of TSS I/O permission bitmap.
* We intercept those #GP and allow access to them anyway
- * as VMware does. Don't intercept #GP for SEV guests as KVM can't
- * decrypt guest memory to decode the faulting instruction.
+ * as VMware does.
*/
- if (enable_vmware_backdoor && !sev_guest(vcpu->kvm))
+ if (enable_vmware_backdoor)
set_exception_intercept(svm, GP_VECTOR);
svm_set_intercept(svm, INTERCEPT_INTR);
@@ -1950,7 +1949,7 @@ static void svm_sync_dirty_debug_regs(struct kvm_vcpu *vcpu)
{
struct vcpu_svm *svm = to_svm(vcpu);
- if (vcpu->arch.guest_state_protected)
+ if (WARN_ON_ONCE(sev_es_guest(vcpu->kvm)))
return;
get_debugreg(vcpu->arch.db[0], 0);
@@ -2681,7 +2680,7 @@ static int dr_interception(struct kvm_vcpu *vcpu)
unsigned long val;
int err = 0;
- if (vcpu->guest_debug == 0) {
+ if (vcpu->guest_debug == 0 && !sev_es_guest(vcpu->kvm)) {
This will change the current flow of an SEV-ES guest. With SEV-ES,
vcpu->guest_debug can never be anything other than 0 and currently always
takes this path.
So what is really needed is:
if (vcpu->guest_debug == 0) {
if (!sev_es_guest(vcpu->kvm)) {
...
}
return 1;
}
/*
* No more DR vmexits; force a reload of the debug registers
* and reenter on this instruction. The next vmexit will
diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
index f44751dd8d5d..7c99a7d55476 100644
--- a/arch/x86/kvm/svm/svm.h
+++ b/arch/x86/kvm/svm/svm.h
@@ -409,23 +409,25 @@ static inline void set_dr_intercepts(struct vcpu_svm *svm)
{
struct vmcb *vmcb = svm->vmcb01.ptr;
- if (!sev_es_guest(svm->vcpu.kvm)) {
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_WRITE);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_WRITE);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_WRITE);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_WRITE);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_WRITE);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_WRITE);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_WRITE);
+ if (sev_es_guest(svm->vcpu.kvm)) {
+ WARN_ON_ONCE(svm->vcpu.arch.last_vmentry_cpu != -1);
+ return;
}
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR0_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR1_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR2_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR3_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR4_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR5_WRITE);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR6_WRITE);
vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
@@ -436,13 +438,13 @@ static inline void clr_dr_intercepts(struct vcpu_svm *svm)
{
struct vmcb *vmcb = svm->vmcb01.ptr;
+ if (WARN_ON_ONCE(sev_es_guest(svm->vcpu.kvm)))
+ return;
+
vmcb->control.intercepts[INTERCEPT_DR] = 0;
- /* DR7 access must remain intercepted for an SEV-ES guest */
- if (sev_es_guest(svm->vcpu.kvm)) {
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
- vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
- }
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_READ);
+ vmcb_set_intercept(&vmcb->control, INTERCEPT_DR7_WRITE);
If we never call clr_dr_intercepts() anymore for an SEV-ES guest, then the
above two lines should be removed. They only were executed for an SEV-ES
guest and now they would be executed for any guest.
Thanks,
Tom
recalc_intercepts(svm);
}
