KVM CPU hotplug notifier triggers BUG_ON on arm64

[Kristina Martsenko]

On 03/07/2023 10:45, Marc Zyngier wrote:
> On Sat, 01 Jul 2023 18:42:28 +0100,
> Oliver Upton <oliver.upton at linux.dev> wrote:
>>
>> Hi Kristina,
>>
>> Thanks for the bug report.
>>
>> On Sat, Jul 01, 2023 at 01:50:52PM +0100, Kristina Martsenko wrote:
>>> Hi,
>>>
>>> When I try to online a CPU on arm64 while a KVM guest is running, I hit a
>>> BUG_ON(preemptible()) (as well as a WARN_ON). See below for the full log.
>>>
>>> This is on kvmarm/next, but seems to have been broken since 6.3. Bisecting it
>>> points at commit:
>>>
>>>   0bf50497f03b ("KVM: Drop kvm_count_lock and instead protect kvm_usage_count with kvm_lock")
>>
>> Makes sense. We were using a spinlock before, which implictly disables
>> preemption.
>>
>> Well, one way to hack around the problem would be to just cram
>> preempt_{disable,enable}() into kvm_arch_hardware_disable(), but that's
>> kinda gross in the context of cpuhp which isn't migratable in the first
>> place. Let me have a look...
> 
> An alternative would be to replace the preemptible() checks with a one
> that looks at the migration state, but I'm not sure that's much better
> (it certainly looks more costly).
> 
> There is also the fact that most of our per-CPU accessors are already
> using preemption disabling, and this code has a bunch of them. So I'm
> not sure there is a lot to be gained from not disabling preemption
> upfront.
> 
> Anyway, as I was able to reproduce the issue under NV, I tested the
> hack below. If anything, I expect it to be a reasonable fix for
> 6.3/6.4, and until we come up with a better approach.
> 
> Thanks,
> 
> 	M.
> 
> diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> index aaeae1145359..a28c4ffe4932 100644
> --- a/arch/arm64/kvm/arm.c
> +++ b/arch/arm64/kvm/arm.c
> @@ -1894,8 +1894,17 @@ static void _kvm_arch_hardware_enable(void *discard)
>  
>  int kvm_arch_hardware_enable(void)
>  {
> -	int was_enabled = __this_cpu_read(kvm_arm_hardware_enabled);
> +	int was_enabled;
>  
> +	/*
> +	 * Most calls to this function are made with migration
> +	 * disabled, but not with preemption disabled. The former is
> +	 * enough to ensure correctness, but most of the helpers
> +	 * expect the later and will throw a tantrum otherwise.
> +	 */
> +	preempt_disable();
> +
> +	was_enabled = __this_cpu_read(kvm_arm_hardware_enabled);
>  	_kvm_arch_hardware_enable(NULL);
>  
>  	if (!was_enabled) {
> @@ -1903,6 +1912,8 @@ int kvm_arch_hardware_enable(void)
>  		kvm_timer_cpu_up();
>  	}
>  
> +	preempt_enable();
> +
>  	return 0;
>  }

This fixes the issue for me.

Tested-by: Kristina Martsenko <kristina.martsenko at arm.com>

Thanks,
Kristina