Re: [PATCH 2/3] KVM: selftests: Introduce kvm_vm_dead_free

[Sean Christopherson <seanjc@google.com>]

On Thu, Nov 07, 2024, Eric Auger wrote:
> In case a KVM_REQ_VM_DEAD request was sent to a VM, subsequent
> KVM ioctls will fail and cause test failure. This now happens
> with an aarch64 vgic test where the kvm_vm_free() fails. Let's
> add a new kvm_vm_dead_free() helper that does all the deallocation
> besides the KVM_SET_USER_MEMORY_REGION2 ioctl.

Please no.  I don't want to bleed the kvm->vm_dead behavior all over selftests.
The hack in __TEST_ASSERT_VM_VCPU_IOCTL() is there purely to provide users with
a more helpful error message, it is most definitely not intended to be an "official"
way to detect and react to the VM being dead.

IMO, tests that intentionally result in a dead VM should assert that subsequent
VM/vCPU ioctls return -EIO, and that's all.  Attempting to gracefully free
resources adds complexity and pollutes the core selftests APIs, with very little
benefit.

Marking a VM dead should be a _very_ rare event; it's not something that I think
we should encourage, i.e. we shouldn't make it easier to deal with.  Ideally,
use of kvm_vm_dead() should be limited to things like sev_vm_move_enc_context_from(),
where KVM needs to prever accessing the source VM to protect the host.  IMO, the
vGIC case and x86's enter_smm() are hacks.  E.g. I don't see any reason why the
enter_smm() case can't synthesize a triple fault.

If memory utilization is a concern for the vgic_init test (which seems unlikely),
I would much rather solve that problem by expanding KVM selftests support for the
selftests harness so that each testcase runs as a child process.

https://lore.kernel.org/all/ZjUwqEXPA5QVItyX@google.com