Oof.  And as Chao pointed out[*], removing the vm_dead check would allow creating
and running vCPUs in a dead VM, which is most definitely not desirable.  Squashing
the vCPU creation case is easy enough if we keep vm_dead but still generally allow
ioctls, and it's probably worth doing that no matter what (to plug the hole where
pending vCPU creations could succeed):

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index d477a7fda0ae..941d2c32b7dc 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c

@@ -4207,6 +4207,11 @@ static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, unsigned long id)

       mutex_lock(&kvm->lock);

+       if (kvm->vm_dead) {
+               r = -EIO;
+               goto unlock_vcpu_destroy;
+       }
+

       if (kvm_get_vcpu_by_id(kvm, id)) {
               r = -EEXIST;
               goto unlock_vcpu_destroy;

And then to ensure vCPUs can't do anything, check KVM_REQ_VM_DEAD after acquiring
vcpu->mutex.

diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 6c07dd423458..883077eee4ce 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c

@@ -4433,6 +4433,12 @@ static long kvm_vcpu_ioctl(struct file *filp,

       if (mutex_lock_killable(&vcpu->mutex))
               return -EINTR;
+
+       if (kvm_test_request(KVM_REQ_VM_DEAD, vcpu)) {
+               r = -EIO;
+               goto out;
+       }
+
       switch (ioctl) {
       case KVM_RUN: {
               struct pid *oldpid;

That should address all TDVPS paths (I hope), and I _think_ would address all
MMU-related paths as well?  E.g. prefault requires a vCPU.

Disallowing (most) vCPU ioctls but not all VM ioctls on vm_dead isn't great ABI
(understatement), but I think we need/want the above changes even if we keep the
general vm_dead restriction.  And given the extremely ad hoc behavior of taking
kvm->lock for VM ioctls, trying to enforce vm_dead for "all" VM ioctls seems like
a fool's errand.

So I'm leaning toward keeping "KVM: Reject ioctls only if the VM is bugged, not
simply marked dead" (with a different shortlog+changelog), but keeping vm_dead
(and not introducing kvm_tdx.vm_terminated).