On 4/15/2022 11:25 PM, Sean Christopherson wrote:
On Mon, Apr 11, 2022, Zeng Guang wrote:
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index d1a39285deab..23fbf52f7bea 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -11180,11 +11180,15 @@ static int sync_regs(struct kvm_vcpu *vcpu)
int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
{
+ int ret = 0;
+
if (kvm_check_tsc_unstable() && atomic_read(&kvm->online_vcpus) != 0)
pr_warn_once("kvm: SMP vm created on host with unstable TSC; "
"guest TSC will not be reliable\n");
- return 0;
+ if (kvm_x86_ops.alloc_ipiv_pid_table)
+ ret = static_call(kvm_x86_alloc_ipiv_pid_table)(kvm);
Add a generic kvm_x86_ops.vcpu_precreate, no reason to make this so specific.
And use KVM_X86_OP_RET0 instead of KVM_X86_OP_OPTIONAL, then this can simply be
return static_call(kvm_x86_vcpu_precreate);
That said, there's a flaw in my genius plan.
1. KVM_CREATE_VM
2. KVM_CAP_MAX_VCPU_ID, set max_vcpu_ids=1
3. KVM_CREATE_VCPU, create IPIv table but ultimately fails
4. KVM decrements created_vcpus back to '0'
5. KVM_CAP_MAX_VCPU_ID, set max_vcpu_ids=4096
6. KVM_CREATE_VCPU w/ ID out of range
In other words, malicious userspace could trigger buffer overflow.
This is the tricky exploit that make max_vcpu_ids update more times. I
think we
can avoid this issue by checking pid table availability during the
pre-creation
of the first vCPU. If it's already allocated, free it and re-allocate to
accommodate
table size to new max_vcpu_ids if updated.
That could be solved by adding an arch hook to undo precreate, but that's gross
and a good indication that we're trying to solve this the wrong way.
I think it's high time we add KVM_FINALIZE_VM, though that's probably a bad name
since e.g. TDX wants to use that name for VM really, really, being finalized[*],
i.e. after all vCPUs have been created.
KVM_POST_CREATE_VM? That's not very good either.
Paolo or anyone else, thoughts?
[*] https://lore.kernel.org/all/83768bf0f786d24f49d9b698a45ba65441ef5ef0.1646422845.git.isaku.yamahata@xxxxxxxxx
