[PATCH 1/2] KVM: arm/arm64: vgic: Don't populate multiple LRs with the same vintid

Christoffer Dall cdall at kernel.org
Fri Mar 9 13:39:31 PST 2018


On Thu, Mar 08, 2018 at 06:39:20PM +0000, Marc Zyngier wrote:
> On Thu, 08 Mar 2018 17:04:38 +0000,
> Marc Zyngier wrote:
> > 
> > On Thu, 08 Mar 2018 16:02:42 +0000,
> > Christoffer Dall wrote:
> > > 
> > > On Thu, Mar 08, 2018 at 10:19:49AM +0000, Marc Zyngier wrote:
> > > > On 07/03/18 23:34, Christoffer Dall wrote:
> > > > > On Wed, Mar 7, 2018 at 12:40 PM, Marc Zyngier <marc.zyngier at arm.com> wrote:
> > > > >> The vgic code is trying to be clever when injecting GICv2 SGIs,
> > > > >> and will happily populate LRs with the same interrupt number if
> > > > >> they come from multiple vcpus (after all, they are distinct
> > > > >> interrupt sources).
> > > > >>
> > > > >> Unfortunately, this is against the letter of the architecture,
> > > > >> and the GICv2 architecture spec says "Each valid interrupt stored
> > > > >> in the List registers must have a unique VirtualID for that
> > > > >> virtual CPU interface.". GICv3 has similar (although slightly
> > > > >> ambiguous) restrictions.
> > > > >>
> > > > >> This results in guests locking up when using GICv2-on-GICv3, for
> > > > >> example. The obvious fix is to stop trying so hard, and inject
> > > > >> a single vcpu per SGI per guest entry. After all, pending SGIs
> > > > >> with multiple source vcpus are pretty rare, and are mostly seen
> > > > >> in scenario where the physical CPUs are severely overcomitted.
> > > > >>
> > > > >> Cc: stable at vger.kernel.org
> > > > >> Fixes: 0919e84c0fc1 ("KVM: arm/arm64: vgic-new: Add IRQ sync/flush framework")
> > > > >> Signed-off-by: Marc Zyngier <marc.zyngier at arm.com>
> > > > >> ---
> > > > >>  virt/kvm/arm/vgic/vgic.c | 11 +----------
> > > > >>  1 file changed, 1 insertion(+), 10 deletions(-)
> > > > >>
> > > > >> diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
> > > > >> index c7c5ef190afa..1f7ff175f47b 100644
> > > > >> --- a/virt/kvm/arm/vgic/vgic.c
> > > > >> +++ b/virt/kvm/arm/vgic/vgic.c
> > > > >> @@ -720,18 +720,9 @@ static void vgic_flush_lr_state(struct kvm_vcpu *vcpu)
> > > > >>         list_for_each_entry(irq, &vgic_cpu->ap_list_head, ap_list) {
> > > > >>                 spin_lock(&irq->irq_lock);
> > > > >>
> > > > >> -               if (unlikely(vgic_target_oracle(irq) != vcpu))
> > > > >> -                       goto next;
> > > > >> -
> > > > >> -               /*
> > > > >> -                * If we get an SGI with multiple sources, try to get
> > > > >> -                * them in all at once.
> > > > >> -                */
> > > > >> -               do {
> > > > >> +               if (likely(vgic_target_oracle(irq) == vcpu))
> > > > >>                         vgic_populate_lr(vcpu, irq, count++);
> > > > > 
> > > > > I think we need to change vgic_populate_lr to set the EOI maintenance
> > > > > interrupt flag so that when the interrupt is deactivated, if there are
> > > > > additional pending sources, we exit the guest and pick up the
> > > > > interrupt.
> > > > 
> > > > Potentially. We need to be careful about about the semantics of EOI MI
> > > > with non-level interrupts (see the other thread about EOI signalling).
> > > 
> > > I'll have a look.
> > > 
> > > > 
> > > > > An alternative would be to set the underflow interrupt, but I don't
> > > > > think that would be correct for multiple priorities, because the SGI
> > > > > could have a higher priority than other pending interrupts we put in
> > > > > the LR.
> > > > 
> > > > I don't think priorities are the issue (after all, we already sort the
> > > > irqs in order of priority). 
> > > 
> > > Yes, but assume you have three pending interrupts, one SGI from two
> > > sources, and one SPI, and assume that the SGI has priority 1 and SPI
> > > priority 2 (lower means higher priority), then I think with underflow or
> > > the no-pending interrupt flag, we'll deliver the SGI from the first
> > > source, and then the SPI, and then exiting to pick up the last SGI from
> > > the other source.  That's not how I understand the GIC architecture is
> > > supposed to work.  Am I missing something?
> > 
> > No, you do have a point here. I'm so glad this is a v2 only thing.
> > 
> > > 
> > > > My worry is that underflow is allowed to
> > > > fire if there is one interrupt pending, which implies that you could
> > > > end-up in a livelock scenario if you only have one SGI pending with
> > > > multiple sources.
> > > 
> > > Yes, doesn't work, so I think it should be a maintenance interrupt on
> > > EOI.
> 
> Thinking of it a bit more: MI on EOI doesn't offer much more guarantee
> in the way of priority ordering. Taking your example above: Even if
> you generate a MI when EOIing the SGI, there is no guarantee that
> you'll take the MI before you've acked the SPI.

There's no guarantee, but at least you're attempting at processing the
SGIs in first.  It's the best we can do, but not completely correct,
kinda thing...

> 
> If you really want to offer that absolute guarantee that all the
> multi-source SGIs of higher priority are delivered before anything
> else, then you must make sure that only the SGIs are present in the
> LRs, excluding any other interrupt on lower priority until you've
> queued them all.

Yes, that sucks!  Might not be too hard to implement, it's basically an
early out of the loop traversing the AP list, but just an annoying
complication.

> 
> At that stage, I wonder if there is a point in doing anything at
> all. The GICv2 architecture is too rubbish for words.
> 

The case we do need to worry about is the guest processing all its
interrupts and not exiting while there is actually still an SGI pending.
At that point, we can either do it with the "no interrupts pending
maintenance interrupt" or with the "EOI maintenance interrupt", and I
think the latter at least gets us slightly closer to the architecture
for a non-virtualized system.

Thanks,
-Christoffer



More information about the linux-arm-kernel mailing list