Re: [PATCH] KVM: Fix error path in kvm_vm_ioctl_create_vcpu() on xa_store() failure

[Sean Christopherson <seanjc@google.com>]

On Wed, Jul 31, 2024, Michal Luczaj wrote:
> On 7/31/24 15:31, Will Deacon wrote:
> > On Tue, Jul 30, 2024 at 04:31:08PM -0700, Sean Christopherson wrote:
> >> On Tue, Jul 30, 2024, Michal Luczaj wrote:
> >>> On 7/30/24 17:56, Will Deacon wrote:
> >>>> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> >>>> index d0788d0a72cc..b80dd8cead8c 100644
> >>>> --- a/virt/kvm/kvm_main.c
> >>>> +++ b/virt/kvm/kvm_main.c
> >>>> @@ -4293,7 +4293,7 @@ static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, unsigned long id)
> >>>>  
> >>>>  	if (KVM_BUG_ON(xa_store(&kvm->vcpu_array, vcpu->vcpu_idx, vcpu, 0), kvm)) {
> >>>>  		r = -EINVAL;
> >>>> -		goto kvm_put_xa_release;
> >>>> +		goto err_xa_release;
> >>>>  	}
> >>>>  
> >>>>  	/*
> >>>> @@ -4310,6 +4310,7 @@ static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, unsigned long id)
> >>>>  
> >>>>  kvm_put_xa_release:
> >>>>  	kvm_put_kvm_no_destroy(kvm);
> >>>> +err_xa_release:
> >>>>  	xa_release(&kvm->vcpu_array, vcpu->vcpu_idx);
> >>>>  unlock_vcpu_destroy:
> >>>>  	mutex_unlock(&kvm->lock);
> >>>
> >>> My bad for neglecting the "impossible" path. Thanks for the fix.
> >>>
> >>> I wonder if it's complete. If we really want to consider the possibility of
> >>> this xa_store() failing, then keeping vCPU fd installed and calling
> >>> kmem_cache_free(kvm_vcpu_cache, vcpu) on the error path looks wrong.
> >>
> >> Yeah, the vCPU is exposed to userspace, freeing its assets will just cause
> >> different problems.  KVM_BUG_ON() will prevent _new_ vCPU ioctl() calls (and kick
> >> running vCPUs out of the guest), but it doesn't interrupt other CPUs, e.g. if
> >> userspace is being sneaking and has already invoked a vCPU ioctl(), KVM will hit
> >> a use-after-free (several of them).
> > 
> > Damn, yes. Just because we haven't returned the fd yet, doesn't mean
> > userspace can't make use of it.
> >
> >> As Michal alluded to, it should be impossible for xa_store() to fail since KVM
> >> pre-allocates/reserves memory.  Given that, deliberately leaking the vCPU seems
> >> like the least awful "solution".
> > 
> > Could we actually just move the xa_store() before the fd creation? I
> > can't immediately see any issues with that...
> 
> Hah, please see commit afb2acb2e3a3 :) Long story short: create_vcpu_fd()
> can legally fail, which must be handled gracefully, which would involve
> destruction of an already xa_store()ed vCPU, which is racy.

Ya, the basic problem is that we have two ways of publishing the vCPU, fd and
vcpu_array, with no way of setting both atomically.  Given that xa_store() should
never fail, I vote we do the simple thing and deliberately leak the memory.