[PATCH 07/10] KVM: arm/arm64: vgic: Allow HW interrupts to be queued to a guest

[Eric Auger]

On 06/11/2015 12:02 PM, Marc Zyngier wrote:
> On 11/06/15 10:44, Andre Przywara wrote:
>> On 06/11/2015 10:15 AM, Marc Zyngier wrote:
>>> On 11/06/15 09:44, Andre Przywara wrote:
>>>> On 06/08/2015 06:04 PM, Marc Zyngier wrote:
>> ...
>>>>> @@ -1344,6 +1364,35 @@ static bool vgic_process_maintenance(struct kvm_vcpu *vcpu)
>>>>>  	return level_pending;
>>>>>  }
>>>>>  
>>>>> +/* Return 1 if HW interrupt went from active to inactive, and 0 otherwise */
>>>>> +static int vgic_sync_hwirq(struct kvm_vcpu *vcpu, struct vgic_lr vlr)
>>>>> +{
>>>>> +	struct irq_phys_map *map;
>>>>> +	int ret;
>>>>> +
>>>>> +	if (!(vlr.state & LR_HW))
>>>>> +		return 0;
>>>>> +
>>>>> +	map = vgic_irq_map_search(vcpu, vlr.irq);
>>>>
>>>> I wonder if it's safe to rely on that mapping here. Are we sure that
>>>> this hasn't changed while the VCPU was running? If I got this correctly,
>>>> currently only vcpu_reset will actually add a map entry, but I guess in
>>>> the future there will be more users.
>>>
>>> How can the guest interrupt change? This is HW, as far as the guest is
>>> concerned. An actual interrupt line. We don't reconfigure the HW live.
>>
>> I was thinking about the rbtree mapping we introduced. There we map a
>> guest interrupt to a hardware interrupt. Are we sure that no one tears
>> down that mapping while we have an LR populated with this pair?
>> I am not talking about the timer here, but more about future users.
>>
>>>> Also we rely on the irqdomain mapping to be still the same, but that is
>>>> probably a safe assumption.
>>>
>>> Like I said before, this *cannot* change.
>>
>> OK, got it.
>>
>>>
>>>> But I'd still find it more natural to use the hwirq number from the LR
>>>> at this point. Can't we use irq_find_mapping() here to learn Linux'
>>>> (current) irq number from that?
>>>
>>> I think you're confused.
>>>
>>> - The guest irq (vlr.irq) is entirely made up, and has no connection
>>> with reality. it is stable, and cannot change during the lifetime of the
>>> guest (think of it as a HW irq line).
>>>
>>> - The host hwirq (vlr.hwirq) is stable as well, for the same reason.
>>>
>>> - The Linux IRQ cannot change because we've been given it by the kernel,
>>> and that's what we use for *everything* as far as the kernel is
>>> concerned. Its mapping to hwirq is stable as well because this is how we
>>> talk to the HW.
>>
>> Not disputing any of them, but:
>>
>>> - irq_find_mapping gives you the *reverse* mapping (from hwirq to Linux
>>> irq), and for that to work, you need the domain on which you want to
>>> apply the translation. This is only useful when actually taking the
>>> interrupt (i.e. in an interrupt controller driver). I can't see how that
>>> could make sense here.
>>
>> So if the guest has acked/EOIed it's IRQ, the GIC at the same time
>> acked/EOIed the hardware IRQ it found in the LR. Now we assume that this
>> is the very same as the HW IRQ we found doing our rbtree traversal.
>> I just wanted to be sure that this is always true and that this mapping
>> didn't change while the VCPU was running.
>> If you are sure of this, fine, I was just concerned that someone breaks
>> this assumption in the future by more dynamically mapping/unmapping
>> entries (say some irq forwarding user) and we will not notice.
> 
> How can the mapping change? Are you thinking of an unmap/map operation
> being done while the guest is running, replacing a HW device with
> another? That's not an option, and not only for the interrupts.

Well that's what we achieved I think with the kvm-vfio integration. The
requirement was: since we allow the user-space to turn forwarding on,
through the kvm-vfio device, we should offer the inverse operation and
this was should never fail. This was achieved by forcing the guest exit,
check the HW state of the IRQ, and quite a lot of pain ...

At that time the kvm-vfio integration seemed to be the most appropriate
approach. Now it seems this is put into question again with Intel posted
IRQ API series review (https://lkml.org/lkml/2015/6/12/595). I think you
will happy. not sure I can say the same ;-)

Best Regards

Eric

> 
> 	M.
>