[RFC] Saving/Restoring the internal state of an interrupt

[Marc Zyngier]

On Mon, Aug 04 2014 at 12:57:16 pm BST, Christoffer Dall <christoffer.dall at linaro.org> wrote:
> On Mon, Jun 16, 2014 at 07:47:57PM +0100, Marc Zyngier wrote:
>> Thomas, all,
>> 
>> Since ARM has grown some virtualization extensions, there is a
>> particular issue I've carefully avoided to address, and which is now
>> biting me exactly where you think it does.
>> 
>> Let's expose the problem:
>> - We have a per-CPU timer (the architected timer)
>> - This timer is shared across virtual machines
>> - We context-switch the state of the timer on entry/exit of the VM
>> 
>> The state of the timer itself is basically a control word and a
>> comparator value. Not much. Ah yes, one tiny detail: the state of the
>> interrupt controller (the GIC) is also part of the state.
>> 
>> This interrupt state is not the pending state, but the internal state
>> (called the "active" state), indicating if the interrupt is in
>> progress or not. If the interrupt is active (and configured as a level
>> interrupt), then the interrupt can't fire again until the guest EOIs
>> it (there is some special hardware dedicated to this).
>
> Can you clarify this part.  I was under the impression from reading the
> GICv2 specs at least that an active interrupt can become active and
> pending (and the pending part of that would cause the interrupt to fire
> again).  Specifically, I'm thinking about transition A2 in Figure 3-1 in
> the GICv2 spec (ARM IHI 0048B.b).  Am I reading this incorrectly or is
> there something special about the timer in this regard?

Nothing special about the timer. Just that the timer being a level
interrupt, the pending state is a direct consequence of the level being
high (as mentionned in the note at the bottom of 1.4.2).

You go from Pending to Active-and-Pending, and then to Active (when the
timer is reprogrammed/disabled, though you may go back to A&P if the timer
fires again immediately). Eventually, after the EOI, you go back to
Inactive.

Here, the only bit we really care about is the Active bit (Pending
changes anyway, as we save/restore the timer context).

Or is there something that I don't understand in your question?

>> 
>> So far, I got away with ignoring that state altogether by playing some
>> tricks in the timer control register (basically masking the interrupt
>> there). This worked fine until Ian found some interesting guest (QNX)
>> that falls over if someone masks the interrupt in the timer behind its
>> back. Fair enough.
>> 
>> So I'm back to square one and I have to context-switch that single
>> bit.
>> 
>> So far, I can see about three options:
>> 
>> - I can add a pair of exported functions:
>>   void gicv2_irq_set_active_state(struct irq_data *d, bool act)
>>   bool gicv2_irq_is_active(struct irq_data *d);
>> 
>>   that directly play with the state. Oh, and for GICv3 as well. And
>>   whatever comes after. It means that KVM has to know exactly the type
>>   of interrupt controller we're using (well, it already does), and call
>>   the right stuff.
>> 
>> - I can add similar function to struct irq_chip:
>>   void (*irq_set_state)(struct irq_data *d, void *data);
>>   void (*irq_get_state)(struct irq_data *d, void *data);
>> 
>>   and build a generic API on top of that. That's tempting, but I'm
>>   really not keen on the "void *data" crap, as it means KVM has to
>>   know the type of the opaque data anyway. Or we define what is the
>>   state of an interrupt, and I'm afraid there is as many definitions
>>   as there are interrupt controllers.
>> 
>> - The third possibility is that I go and duplicate parts of the two GIC
>>   drivers into KVM just to be able to save/restore these bits. Please
>>   pass the bucket around.
>> 
>> So far, I've prototyped the first option, but I'm seriously
>> questioning my own sanity. Any idea/opinion?
>> 
> KVM and the GIC driver are tightly bound anyhow, so is there anything
> particular horrible in the first approach?  (Besides the fact that we're
> exposing random bits of information to the entire kernel not intended to
> be used by anyone else than KVM for now).
>
> Second approach feels over-engineered to me, unless there are other
> needs or real use cases in the near term.

Well, I've decided on option 2 so far, as I feel we'd better have
something that is more generic then just a bunch of driver-specific
hacks. Other architectures may have similar requirements (we just
haven't realized it yet).

> The third option, just feels wrong, we'd have to do lookups in the DT
> again to get the base address of the GICC etc. right?

Yup. I don't even want to think about it. ;-)

	M.
-- 
Jazz is not dead. It just smells funny.