Re: [PATCH] KVM: x86: Allocate/free user_return_msrs at kvm.ko (un)loading time

[Sean Christopherson]

On Mon, Nov 10, 2025, Chao Gao wrote:
> >-static int kvm_init_user_return_msrs(void)
> >+static void kvm_destroy_user_return_msrs(void)
> > {
> >-	user_return_msrs = alloc_percpu(struct kvm_user_return_msrs);
> >-	if (!user_return_msrs) {
> >-		pr_err("failed to allocate percpu user_return_msrs\n");
> >-		return -ENOMEM;
> >-	}
> >+	int cpu;
> >+
> >+	for_each_possible_cpu(cpu)
> >+		WARN_ON_ONCE(per_cpu(user_return_msrs, cpu).registered);
> 
> Could this warning be triggered if the forced shutdown path didn't
> unregister the user return callback (i.e., with the patch [*] applied),
> and then vendor modules got unloaded immediately after the forced shutdown
> (before the CPU exits to the userspace)?

Probably?  But that's more of a feature than a bug, e.g. gives the (privileged!)
user the heads up of how exactly they broke their system when they forced a
reboot.  I'd prefer not to condition it on e.g. !kvm_rebooting unless it's truly
necessary, because it's "just" a WARN, i.e. shouldn't crash the system.

> [*]: https://lore.kernel.org/kvm/20251030191528.3380553-4-seanjc@xxxxxxxxxx/