Skip to contentSkip to conversation

KVM: nSVM: Check full nested state size before writing to user buffer

PATCH Superseded Priyanka Mani 4 participants 5 messages Nov 12

Subsystems (3)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KVM: nSVM: Check full nested state size before writing to user buffer

svm_get_nested_state() compares the userspace-provided buffer size against kvm_state.size before adding KVM_STATE_NESTED_SVM_VMCB_SIZE for a vCPU that is in guest mode. When L2 is active and userspace passes a size that is large enough for the header but too small for the vmcb12 payload, KVM thus skips the early exit, copies out the header, clears and fills the full 4KiB vmcb12 area past the end of the buffer described by userspace, and only then does the caller notice that the returned size is larger than the provided one and fail with -E2BIG.

As a result, KVM writes up to 4KiB beyond the buffer userspace said it has, and if that memory isn't mapped, the ioctl fails with -EFAULT instead of the documented -E2BIG without reporting the required size.

Compute the full size before the check, as vmx_get_nested_state() does, so that a too-small buffer is rejected without writing anything to it.

Fixes cc440cdad5b7 ("KVM: nSVM: implement KVM_GET_NESTED_STATE and KVM_SET_NESTED_STATE")

---
Tested on an AMD EPYC 7763 host by booting the patched kernel under QEMU (nested SVM) with KASAN and lockdep. A reproducer enters L2, exits to userspace from L2 and calls KVM_GET_NESTED_STATE with
size = sizeof(struct kvm_nested_state). Without this patch, KVM writes 4096 bytes past the 128-byte buffer before returning -E2BIG, or returns -EFAULT if the next page is unmapped. With this patch, KVM returns -E2BIG and reports size = 4224 without writing to the buffer.

KVM selftests nested_set_state_test, state_test, smm_test, hyperv_svm_test, svm_nested_vmcb12_gpa, svm_nested_pat_test and the other nested SVM tests give identical results with and without the patch. kvm-unit-tests' SVM tests could not be run in this nested setup. Not tested on Intel; the VMX path is unchanged.

 arch/x86/kvm/svm/nested.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
arch/x86/kvm/svm/nested.c +3 -1
index 1cecb82702e2..9a70136172f3 100644
@@ -1916,6 +1916,9 @@ static int svm_get_nested_state(struct kvm_vcpu *vcpu,
 
 	svm = to_svm(vcpu);
 
+	if (is_guest_mode(vcpu))
+		kvm_state.size += KVM_STATE_NESTED_SVM_VMCB_SIZE;
+
 	if (user_data_size < kvm_state.size)
 		goto out;
 
@@ -1925,7 +1928,6 @@ static int svm_get_nested_state(struct kvm_vcpu *vcpu,
 		kvm_state.hdr.svm.gpat = 0;
 		if (l2_has_separate_pat(vcpu))
 			kvm_state.hdr.svm.gpat = svm->vmcb->save.g_pat;
-		kvm_state.size += KVM_STATE_NESTED_SVM_VMCB_SIZE;
 		kvm_state.flags |= KVM_STATE_NESTED_GUEST_MODE;
 
 		if (vcpu->arch.nested_run_pending)

base-commit: cbe54fc3462b3949ab27b9ae81fb9ddb509040c7

Signature · 1 line
2.43.0
#1 lore raw
Thank you for your contribution! Sashiko AI review found no new issues in this patch, but noticed 1 related pre-existing issue it does not introduce: - [High] KVM: SVM: guest mode and MMU leak in…

Thank you for your contribution! Sashiko AI review found no new issues in this patch, but noticed 1 related pre-existing issue it does not introduce:
- [High] KVM: SVM: guest mode and MMU leak in svm_set_nested_state() [1] --

commit 69ab785aa5333f3d6749b1f4b937158e7dfed5c6
Author: Priyanka Mani <[email protected]>

KVM: nSVM: Check full nested state size before writing to user buffer

This commit adds KVM_STATE_NESTED_SVM_VMCB_SIZE to kvm_state.size before checking user_data_size in svm_get_nested_state() when the vCPU is in guest mode, preventing writes past the end of a too-small userspace buffer.

Quoted 15 lines
arch/x86/kvm/svm/nested.c +3 -0
index 1cecb82702e2c..9a70136172f38 100644
@@ -1916,6 +1916,9 @@ static int svm_get_nested_state(struct kvm_vcpu *vcpu,
 
 	svm = to_svm(vcpu);
 
+	if (is_guest_mode(vcpu))
+		kvm_state.size += KVM_STATE_NESTED_SVM_VMCB_SIZE;
+
 	if (user_data_size < kvm_state.size)
 		goto out;
 

arch/x86/kvm/svm/nested.c static int svm_get_nested_state(struct kvm_vcpu *vcpu, · line 1933

Quoted 7 lines
arch/x86/kvm/svm/nested.c +0 -1
@@ -1925,7 +1928,6 @@ static int svm_get_nested_state(struct kvm_vcpu *vcpu,
 		kvm_state.hdr.svm.gpat = 0;
 		if (l2_has_separate_pat(vcpu))
 			kvm_state.hdr.svm.gpat = svm->vmcb->save.g_pat;
-		kvm_state.size += KVM_STATE_NESTED_SVM_VMCB_SIZE;
 		kvm_state.flags |= KVM_STATE_NESTED_GUEST_MODE;
 
 		if (vcpu->arch.nested_run_pending)

[Severity: High]
This problem wasn't introduced by this patch, but in the paired setter svm_set_nested_state() [1], does a failed nested_svm_load_cr3() leave the vCPU in guest mode with an active nested MMU context?

arch/x86/kvm/svm/nested.c:svm_set_nested_state() {
    ...
	nested_vmcb02_prepare_control(svm);
    ...
	ret = nested_svm_load_cr3(&svm->vcpu, vcpu->arch.cr3,
				  nested_npt_enabled(svm), false);
	if (ret)
		goto out_free;
    ...
}

nested_vmcb02_prepare_control() enters guest mode and initializes the nested MMU context, but the out_free error path only frees save and ctl without calling leave_guest_mode() or nested_svm_uninit_mmu_context(). If userspace passes an invalid guest CR3 and then continues running the vCPU after KVM_SET_NESTED_STATE fails, the vCPU runs with a corrupted nested state and leaked MMU state.

[1] https://sashiko.dev/#/bug/linux-2c0d7c2c-169f-45d3-9953-b6aaf919e66b

Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/57324d544da1e92cf3b30ff7c40595dfbca1933d.1794465097.git.priyankamani2100@gmail.com?part=1
#2 lore raw
Quoted 16 lines · On Thu, Nov 12, 2026, [email protected] wrote:

[Severity: High]
This problem wasn't introduced by this patch, but in the paired setter svm_set_nested_state() [1], does a failed nested_svm_load_cr3() leave the vCPU in guest mode with an active nested MMU context?

arch/x86/kvm/svm/nested.c:svm_set_nested_state() {
    ...
	nested_vmcb02_prepare_control(svm);
    ...
	ret = nested_svm_load_cr3(&svm->vcpu, vcpu->arch.cr3,
				  nested_npt_enabled(svm), false);
	if (ret)
		goto out_free;
    ...
}
Quoted 6 lines

nested_vmcb02_prepare_control() enters guest mode and initializes the nested MMU context, but the out_free error path only frees save and ctl without calling leave_guest_mode() or nested_svm_uninit_mmu_context(). If userspace passes an invalid guest CR3 and then continues running the vCPU after KVM_SET_NESTED_STATE fails, the vCPU runs with a corrupted nested state and leaked MMU state.

Ya, I have this locally, just need to write a changelog and get it posted:

x86/kvm/svm/nested.c +3 -1
index 07554089ec05..4b1cdc7798bb 100644
@@ -2114,8 +2114,10 @@ static int svm_set_nested_state(struct kvm_vcpu *vcpu,
         */
        ret = nested_svm_load_cr3(&svm->vcpu, vcpu->arch.cr3,
                                  nested_npt_enabled(svm), false);
-       if (ret)
+       if (ret) {
+               svm_leave_nested(vcpu);
                goto out_free;
+       }
 
        svm->nested.force_msr_bitmap_recalc = true;
#3 lore raw
In reply to Priyanka Mani
Quoted 9 lines · Priyanka Mani

svm_get_nested_state() compares the userspace-provided buffer size against kvm_state.size before adding KVM_STATE_NESTED_SVM_VMCB_SIZE for a vCPU that is in guest mode. When L2 is active and userspace passes a size that is large enough for the header but too small for the vmcb12 payload, KVM thus skips the early exit, copies out the header, clears and fills the full 4KiB vmcb12 area past the end of the buffer described by userspace, and only then does the caller notice that the returned size is larger than the provided one and fail with -E2BIG.

Quoted 6 lines

As a result, KVM writes up to 4KiB beyond the buffer userspace said it has, and if that memory isn't mapped, the ioctl fails with -EFAULT instead of the documented -E2BIG without reporting the required size.

Compute the full size before the check, as vmx_get_nested_state() does, so that a too-small buffer is rejected without writing anything to it.

I would prefer to instead move the if (user_data_size < kvm_state.size) check until right before the copy_to_user(). It's more future proof that way.

regards,
dan carpenter

#4 lore raw
Quoted 14 lines · Dan Carpenter
Quoted 13 lines · Priyanka Mani

svm_get_nested_state() compares the userspace-provided buffer size against kvm_state.size before adding KVM_STATE_NESTED_SVM_VMCB_SIZE for a vCPU that is in guest mode. When L2 is active and userspace passes a size that is large enough for the header but too small for the vmcb12 payload, KVM thus skips the early exit, copies out the header, clears and fills the full 4KiB vmcb12 area past the end of the buffer described by userspace, and only then does the caller notice that the returned size is larger than the provided one and fail with -E2BIG.

As a result, KVM writes up to 4KiB beyond the buffer userspace said it has, and if that memory isn't mapped, the ioctl fails with -EFAULT instead of the documented -E2BIG without reporting the required size.

Quoted 6 lines
Quoted 2 lines

Compute the full size before the check, as vmx_get_nested_state() does, so that a too-small buffer is rejected without writing anything to it.

I would prefer to instead move the if (user_data_size < kvm_state.size) check until right before the copy_to_user(). It's more future proof that way.

+1

Side topic, your mail got flagged by gmail with:

This message isn't authenticated and the sender can't be verified.