[PATCH v4 07/11] KVM: arm/arm64: vgic: Allow HW interrupts to be queued to a guest

[christoffer.dall@linaro.org (Christoffer Dall)]

On Wed, Sep 30, 2015 at 03:06:32PM +0100, Andre Przywara wrote:
> Hi Christoffer,
> 
> On 29/09/15 14:44, Christoffer Dall wrote:
> > On Wed, Sep 23, 2015 at 06:55:04PM +0100, Andre Przywara wrote:
> >> Salut Marc,
> >>
> >> I know that this patch is already merged, but ....
> >>
> >> On 07/08/15 16:45, Marc Zyngier wrote:
> >>> diff --git a/virt/kvm/arm/vgic.c b/virt/kvm/arm/vgic.c
> >>> index 51c9900..9d009d2 100644
> >> ...
> >>> @@ -1364,6 +1397,39 @@ static bool vgic_process_maintenance(struct kvm_vcpu *vcpu)
> >>>  	return level_pending;
> >>>  }
> >>>  
> >>> +/*
> >>> + * Save the physical active state, and reset it to inactive.
> >>> + *
> >>> + * 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);
> >>> +	BUG_ON(!map || !map->active);
> >>> +
> >>> +	ret = irq_get_irqchip_state(map->irq,
> >>> +				    IRQCHIP_STATE_ACTIVE,
> >>> +				    &map->active);
> >>> +
> >>> +	WARN_ON(ret);
> >>> +
> >>> +	if (map->active) {
> >>> +		ret = irq_set_irqchip_state(map->irq,
> >>> +					    IRQCHIP_STATE_ACTIVE,
> >>> +					    false);
> >>> +		WARN_ON(ret);
> >>> +		return 0;
> >>> +	}
> >>> +
> >>> +	return 1;
> >>> +}
> >>> +
> >>>  /* Sync back the VGIC state after a guest run */
> >>>  static void __kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
> >>>  {
> >>> @@ -1378,14 +1444,31 @@ static void __kvm_vgic_sync_hwstate(struct kvm_vcpu *vcpu)
> >>>  	elrsr = vgic_get_elrsr(vcpu);
> >>>  	elrsr_ptr = u64_to_bitmask(&elrsr);
> >>>  
> >>> -	/* Clear mappings for empty LRs */
> >>> -	for_each_set_bit(lr, elrsr_ptr, vgic->nr_lr) {
> >>> +	/* Deal with HW interrupts, and clear mappings for empty LRs */
> >>> +	for (lr = 0; lr < vgic->nr_lr; lr++) {
> >>>  		struct vgic_lr vlr;
> >>>  
> >>> -		if (!test_and_clear_bit(lr, vgic_cpu->lr_used))
> >>> +		if (!test_bit(lr, vgic_cpu->lr_used))
> >>>  			continue;
> >>>  
> >>>  		vlr = vgic_get_lr(vcpu, lr);
> >>> +		if (vgic_sync_hwirq(vcpu, vlr)) {
> >>> +			/*
> >>> +			 * So this is a HW interrupt that the guest
> >>> +			 * EOI-ed. Clean the LR state and allow the
> >>> +			 * interrupt to be sampled again.
> >>> +			 */
> >>> +			vlr.state = 0;
> >>> +			vlr.hwirq = 0;
> >>> +			vgic_set_lr(vcpu, lr, vlr);
> >>> +			vgic_irq_clear_queued(vcpu, vlr.irq);
> >>
> >> Isn't this line altering common VGIC state without holding the lock?
> >> Eric removed the coarse dist->lock around the whole
> >> __kvm_vgic_sync_hwstate() function, we take it now in
> >> vgic_process_maintenance(), but don't hold it here AFAICT.
> >> As long as we are only dealing with private timer IRQs this is probably
> >> not a problem, but the IRQ number could be a SPI as well, right?
> >>
> > I don't see a problematic race with this though, as all we're doing is
> > to clear a bit in a bitmap, which is always checked atomically, so
> > adding a lock around this really doesn't change anything as far as I can
> > tell.
> 
> Indeed I found a similar comment in some older revisions of the code.
> 
> But isn't it that other code holding the lock (thinking about
> kvm_vgic_flush_hwstate() in particular) assumes that no-one else tinkers
> with the VGIC state while it holds the lock?
> So couldn't we (potentially) run into inconsistent state because we
> cleared the queued bit while the flushing code runs over all interrupts?
> Maybe not in this particular case, but in general?

In general, yes, you should lock operations accessing the distributor.

It just feels silly to do 

spin_lock();
vgic_irq_clear_queued(...);
spin_unlock();

because vgic_irq_clear_queued just clears a bit and I don't see the
race.

> 
> Haven't looked at your new series yet, but will do this ASAP.
> 
Thanks, much appreciated.  Based on your comment on the previous
version, the whole thing is now wrapped in a spinlock so this point is
moot.  Unless there's a clear race to be fixed here, I would prefer if
we focus our energy on getting the other series merged.

-Christoffer