KVM: x86: Allow restore of some sregs with protected state
PATCH 1/1 KVM: x86: Allow restore of some sregs with protected state
With protected state (like SEV-ES and SEV-SNP), KVM does not have direct access to guest registers. However, we deflect modifications to CR0, CR4 and EFER to the host. We also carry the apic_base register and learn about CR8 directly from a VMCB field.
That means these bits of information do exist in the host's KVM data structures. If we ever want to resume consumption of an already initialized VMSA (guest state), we will need to also restore these additional bits of KVM state.
Prepare ourselves for such a world by allowing set_sregs to set the relevant fields. This way, it mirrors get_sregs properly that already exposes them to user space.
--- arch/x86/kvm/x86.c | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)
arch/x86/kvm/x86.c +14 -2
@@ -11370,7 +11370,8 @@ static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs,int idx;struct desc_ptr dt; -if (!kvm_is_valid_sregs (vcpu, sregs)) +if (!vcpu->arch.guest_state_protected && + !kvm_is_valid_sregs (vcpu, sregs))return -EINVAL ; apic_base_msr.data = sregs->apic_base;
@@ -11378,8 +11379,19 @@ static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs,if (kvm_set_apic_base (vcpu, &apic_base_msr))return -EINVAL ; -if (vcpu->arch.guest_state_protected) +if (vcpu->arch.guest_state_protected) { +/* +* While most actual guest state is hidden from us, +* CR{0,4,8}, efer and apic_base still hold KVM state +* with protection enabled, so let's allow restoring +*/ +kvm_set_cr8 (vcpu, sregs->cr8); + kvm_x86_ops.set_efer (vcpu, sregs->efer); + kvm_x86_ops.set_cr0 (vcpu, sregs->cr0); + vcpu->arch.cr0 = sregs->cr0; + kvm_x86_ops.set_cr4 (vcpu, sregs->cr4);return 0 ; + } dt.size = sregs->idt.limit; dt.address = sregs->idt.base;
Signature · 12 lines
2.39.2 Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
Sean Christopherson
maintainer+Peter
Quoted 2 lines · Alexander Graf
With protected state (like SEV-ES and SEV-SNP), KVM does not have direct access to guest registers. However, we deflect modifications to CR0,
Please avoid pronouns in changelogs and comments.
Quoted 7 lines
CR4 and EFER to the host. We also carry the
apic_baseregister and learn about CR8 directly from a VMCB field.That means these bits of information do exist in the host's KVM data structures. If we ever want to resume consumption of an already initialized VMSA (guest state), we will need to also restore these additional bits of KVM state.
For some definitions of "need". I've looked at this code multiple times in the past, and even posted patches[1], but I'm still unconvinced that trapping CR0, CR4, and EFER updates is necessary[2], which is partly why series to fix this stalled out.
: If KVM chugs along happily without these patches, I'd love to pivot and yank out : all of the CR0/4/8 and EFER trapping/tracking, and then make KVM_GET_SREGS a nop : as well.
[1] https://lore.kernel.org/all/[email protected]
[2] https://lore.kernel.org/all/[email protected]
Quoted 14 lines
Prepare ourselves for such a world by allowing
set_sregsto set the relevant fields. This way, it mirrorsget_sregsproperly that already exposes them to user space.Signed-off-by Alexander Graf <[email protected]>--- arch/x86/kvm/x86.c | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)arch/x86/kvm/x86.c +0
-0
arch/x86/kvm/x86.c static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs, · line 11374
Quoted 6 lines
arch/x86/kvm/x86.c +2
-1@@ -11370,7 +11370,8 @@ static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs,int idx;struct desc_ptr dt; -if (!kvm_is_valid_sregs (vcpu, sregs)) +if (!vcpu->arch.guest_state_protected && + !kvm_is_valid_sregs (vcpu, sregs))
This is broken, userspace shouldn't be allowed to stuff complete garbage just because guest state is protected.
Quoted 7 lines
arch/x86/kvm/x86.c +0
-0return -EINVAL ; apic_base_msr.data = sregs->apic_base;@@ -11378,8 +11379,19 @@ static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs,if (kvm_set_apic_base (vcpu, &apic_base_msr))return -EINVAL ;
arch/x86/kvm/x86.c static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs, · line 11390
Quoted 10 lines
arch/x86/kvm/x86.c +9
-1-if (vcpu->arch.guest_state_protected) +if (vcpu->arch.guest_state_protected) { +/* +* While most actual guest state is hidden from us, +* CR{0,4,8}, efer and apic_base still hold KVM state +* with protection enabled, so let's allow restoring +*/ +kvm_set_cr8 (vcpu, sregs->cr8); + kvm_x86_ops.set_efer (vcpu, sregs->efer); + kvm_x86_ops.set_cr0 (vcpu, sregs->cr0);
Use static_call(). This code is also broken in that it doesn't set mmu_reset_needed. This patch also misses the problematic behavior in svm_set_cr0().
And I don't like having a completely separate one-off flow for protected guests. It's more code and arguably uglier, but I would prefer to explicitly skip each chunk as needed so that we aren't effectively maintaining multiple flows.
Easiest thing is probably to just resurrect my aforementioned series, pending the result of the discussion on whether or not KVM should be trapping CR0/CR4/EFER writes in the first place.
Quoted 22 lines · Sean Christopherson
+Peter
Quoted 2 lines · Alexander Graf
With protected state (like SEV-ES and SEV-SNP), KVM does not have direct access to guest registers. However, we deflect modifications to CR0,
Please avoid pronouns in changelogs and comments.
Quoted 7 lines
CR4 and EFER to the host. We also carry the
apic_baseregister and learn about CR8 directly from a VMCB field.That means these bits of information do exist in the host's KVM data structures. If we ever want to resume consumption of an already initialized VMSA (guest state), we will need to also restore these additional bits of KVM state.
For some definitions of "need". I've looked at this code multiple times in the past, and even posted patches[1], but I'm still unconvinced that trapping CR0, CR4, and EFER updates is necessary[2], which is partly why series to fix this stalled out.
Quoted 6 lines
: If KVM chugs along happily without these patches, I'd love to pivot and yank out : all of the CR0/4/8 and EFER trapping/tracking, and then make
KVM_GET_SREGSa nop : as well.[1] https://lore.kernel.org/all/[email protected]
[2] https://lore.kernel.org/all/[email protected]
Yea we are using similar patches to do intra-host migration for SNP VMs.
I have dropped the ball on my AI from that thread. Let me look/test this patch.
Quoted 18 lines
Quoted 18 lines
Prepare ourselves for such a world by allowing
set_sregsto set the relevant fields. This way, it mirrorsget_sregsproperly that already exposes them to user space.Signed-off-by Alexander Graf <[email protected]>--- arch/x86/kvm/x86.c | 16 ++++++++++++++-- 1 file changed, 14 insertions(+), 2 deletions(-)arch/x86/kvm/x86.c +0
-0@@ -11370,7 +11370,8 @@ static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs,int idx;struct desc_ptr dt;
Quoted 6 lines
arch/x86/kvm/x86.c static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs, · line 11374
Quoted 3 lines
arch/x86/kvm/x86.c +2
-1-if (!kvm_is_valid_sregs (vcpu, sregs)) +if (!vcpu->arch.guest_state_protected && + !kvm_is_valid_sregs (vcpu, sregs))This is broken, userspace shouldn't be allowed to stuff complete garbage just because guest state is protected.
Agreed, we've been using something like this:
kvm_valid_sregs()
- if (sregs->fer & EFER_LME && !vcpu->arch.guest_state_protected) {
+ if (sregs->efer & EFER_LME) {
Quoted 29 lines
Quoted 8 lines
arch/x86/kvm/x86.c +0
-0return -EINVAL ; apic_base_msr.data = sregs->apic_base;@@ -11378,8 +11379,19 @@ static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs,if (kvm_set_apic_base (vcpu, &apic_base_msr))return -EINVAL ;arch/x86/kvm/x86.c static int __set_sregs_common(struct kvm_vcpu *vcpu, struct kvm_sregs *sregs, · line 11390
Quoted 10 lines
arch/x86/kvm/x86.c +9
-1-if (vcpu->arch.guest_state_protected) +if (vcpu->arch.guest_state_protected) { +/* +* While most actual guest state is hidden from us, +* CR{0,4,8}, efer and apic_base still hold KVM state +* with protection enabled, so let's allow restoring +*/ +kvm_set_cr8 (vcpu, sregs->cr8); + kvm_x86_ops.set_efer (vcpu, sregs->efer); + kvm_x86_ops.set_cr0 (vcpu, sregs->cr0);Use
static_call(). This code is also broken in that it doesn't setmmu_reset_needed. This patch also misses the problematic behavior insvm_set_cr0().And I don't like having a completely separate one-off flow for protected guests. It's more code and arguably uglier, but I would prefer to explicitly skip each chunk as needed so that we aren't effectively maintaining multiple flows.
Easiest thing is probably to just resurrect my aforementioned series, pending the result of the discussion on whether or not KVM should be trapping CR0/CR4/EFER writes in the first place.
Quoted 24 lines · Peter Gonda
Quoted 23 lines · Sean Christopherson
+Peter
Quoted 2 lines · Alexander Graf
With protected state (like SEV-ES and SEV-SNP), KVM does not have direct access to guest registers. However, we deflect modifications to CR0,
Please avoid pronouns in changelogs and comments.
Quoted 7 lines
CR4 and EFER to the host. We also carry the
apic_baseregister and learn about CR8 directly from a VMCB field.That means these bits of information do exist in the host's KVM data structures. If we ever want to resume consumption of an already initialized VMSA (guest state), we will need to also restore these additional bits of KVM state.
For some definitions of "need". I've looked at this code multiple times in the past, and even posted patches[1], but I'm still unconvinced that trapping CR0, CR4, and EFER updates is necessary[2], which is partly why series to fix this stalled out.
: If KVM chugs along happily without these patches, I'd love to pivot and yank out : all of the CR0/4/8 and EFER trapping/tracking, and then make
KVM_GET_SREGSa nop : as well.
Quoted 5 lines
Quoted 2 lines
[1] https://lore.kernel.org/all/[email protected]
[2] https://lore.kernel.org/all/[email protected]Yea we are using similar patches to do intra-host migration for SNP VMs.
I have dropped the ball on my AI from that thread. Let me look/test this patch.
Awesome, thanks. If we can get away without any of the above states and make sregs completely useless for protected state, I'd be even happier :)
Alex
Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879