On Thu, 14 Nov 2019 11:16:55 +0000
Marc Zyngier <maz@xxxxxxxxxx> wrote:
Hi Marc,
> On 2019-11-12 09:36, Andre Przywara wrote:
> > On Sun, 10 Nov 2019 14:29:14 +0000
> > Marc Zyngier <maz@xxxxxxxxxx> wrote:
> >
> > Hi Marc,
> >
> >> On Fri, 8 Nov 2019 17:49:51 +0000
> >> Andre Przywara <andre.przywara@xxxxxxx> wrote:
> >>
> >> > Our current VGIC emulation code treats the "EnableGrpX" bits in
> >> GICD_CTLR
> >> > as a single global interrupt delivery switch, where in fact the
> >> GIC
> >> > architecture asks for this being separate for the two interrupt
> >> groups.
> >> >
> >> > To implement this properly, we have to slightly adjust our design,
> >> to
> >> > *not* let IRQs from a disabled interrupt group be added to the
> >> ap_list.
> >> >
> >> > As a consequence, enabling one group requires us to re-evaluate
> >> every
> >> > pending IRQ and potentially add it to its respective ap_list.
> >> Similarly
> >> > disabling an interrupt group requires pending IRQs to be removed
> >> from
> >> > the ap_list (as long as they have not been activated yet).
> >> >
> >> > Implement a rather simple, yet not terribly efficient algorithm to
> >> > achieve this: For each VCPU we iterate over all IRQs, checking for
> >> > pending ones and adding them to the list. We hold the ap_list_lock
> >> > for this, to make this atomic from a VCPU's point of view.
> >> >
> >> > When an interrupt group gets disabled, we can't directly remove
> >> affected
> >> > IRQs from the ap_list, as a running VCPU might have already
> >> activated
> >> > them, which wouldn't be immediately visible to the host.
> >> > Instead simply kick all VCPUs, so that they clean their ap_list's
> >> > automatically when running vgic_prune_ap_list().
> >> >
> >> > Signed-off-by: Andre Przywara <andre.przywara@xxxxxxx>
> >> > ---
> >> > virt/kvm/arm/vgic/vgic.c | 88
> >> ++++++++++++++++++++++++++++++++++++----
> >> > 1 file changed, 80 insertions(+), 8 deletions(-)
> >> >
> >> > diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
> >> > index 3b88e14d239f..28d9ff282017 100644
> >> > --- a/virt/kvm/arm/vgic/vgic.c
> >> > +++ b/virt/kvm/arm/vgic/vgic.c
> >> > @@ -339,6 +339,38 @@ int vgic_dist_enable_group(struct kvm *kvm,
> >> int group, bool status)
> >> > return 0;
> >> > }
> >> >
> >> > +/*
> >> > + * Check whether a given IRQs need to be queued to this ap_list,
> >> and do
> >> > + * so if that's the case.
> >> > + * Requires the ap_list_lock to be held (but not the irq lock).
> >> > + *
> >> > + * Returns 1 if that IRQ has been added to the ap_list, and 0 if
> >> not.
> >> > + */
> >> > +static int queue_enabled_irq(struct kvm *kvm, struct kvm_vcpu
> >> *vcpu,
> >> > + int intid)
> >>
> >> true/false seems better than 1/0.
> >
> > Mmh, indeed. I think I had more in there in an earlier version.
> >
> >> > +{
> >> > + struct vgic_irq *irq = vgic_get_irq(kvm, vcpu, intid);
> >> > + int ret = 0;
> >> > +
> >> > + raw_spin_lock(&irq->irq_lock);
> >> > + if (!irq->vcpu && vcpu == vgic_target_oracle(irq)) {
> >> > + /*
> >> > + * Grab a reference to the irq to reflect the
> >> > + * fact that it is now in the ap_list.
> >> > + */
> >> > + vgic_get_irq_kref(irq);
> >> > + list_add_tail(&irq->ap_list,
> >> > + &vcpu->arch.vgic_cpu.ap_list_head);
> >>
> >> Two things:
> >> - This should be the job of vgic_queue_irq_unlock. Why are you
> >> open-coding it?
> >
> > I was *really* keen on reusing that, but couldn't for two reasons:
> > a) the locking code inside vgic_queue_irq_unlock spoils that: It
> > requires the irq_lock to be held, but not the ap_list_lock. Then it
> > takes both locks, but returns with both of them dropped. We need to
> > hold the ap_list_lock all of the time, to prevent any VCPU returning
> > to the HV to interfere with this routine.
> > b) vgic_queue_irq_unlock() kicks the VCPU already, where I want to
> > just add all of them first, then kick the VCPU at the end.
>
> Indeed, and that is why you need to change the way you queue these
> pending, enabled, group-disabled interrupts (see the LPI issue below).
>
> >
> > So I decided to go with the stripped-down version of it, because I
> > didn't dare to touch the original function. I could refactor this
> > "actually add to the list" part of vgic_queue_irq_unlock() into this
> > new function, then call it from both vgic_queue_irq_unlock() and from
> > the new users.
> >
> >> - What if the interrupt isn't pending? Non-pending, non-active
> >> interrupts should not be on the AP list!
> >
> > That should be covered by vgic_target_oracle() already, shouldn't it?
>
> Ah, yes, you're right.
>
> >
> >> > + irq->vcpu = vcpu;
> >> > +
> >> > + ret = 1;
> >> > + }
> >> > + raw_spin_unlock(&irq->irq_lock);
> >> > + vgic_put_irq(kvm, irq);
> >> > +
> >> > + return ret;
> >> > +}
> >> > +
> >> > /*
> >> > * The group enable status of at least one of the groups has
> >> changed.
> >> > * If enabled is true, at least one of the groups got enabled.
> >> > @@ -346,17 +378,57 @@ int vgic_dist_enable_group(struct kvm *kvm,
> >> int group, bool status)
> >> > */
> >> > void vgic_rescan_pending_irqs(struct kvm *kvm, bool enabled)
> >> > {
> >> > + int cpuid;
> >> > + struct kvm_vcpu *vcpu;
> >> > +
> >> > /*
> >> > - * TODO: actually scan *all* IRQs of the VM for pending IRQs.
> >> > - * If a pending IRQ's group is now enabled, add it to its
> >> ap_list.
> >> > - * If a pending IRQ's group is now disabled, kick the VCPU to
> >> > - * let it remove this IRQ from its ap_list. We have to let the
> >> > - * VCPU do it itself, because we can't know the exact state of
> >> an
> >> > - * IRQ pending on a running VCPU.
> >> > + * If no group got enabled, we only have to potentially remove
> >> > + * interrupts from ap_lists. We can't do this here, because a
> >> running
> >> > + * VCPU might have ACKed an IRQ already, which wouldn't
> >> immediately
> >> > + * be reflected in the ap_list.
> >> > + * So kick all VCPUs, which will let them re-evaluate their
> >> ap_lists
> >> > + * by running vgic_prune_ap_list(), removing no longer enabled
> >> > + * IRQs.
> >> > + */
> >> > + if (!enabled) {
> >> > + vgic_kick_vcpus(kvm);
> >> > +
> >> > + return;
> >> > + }
> >> > +
> >> > + /*
> >> > + * At least one group went from disabled to enabled. Now we need
> >> > + * to scan *all* IRQs of the VM for newly group-enabled IRQs.
> >> > + * If a pending IRQ's group is now enabled, add it to the
> >> ap_list.
> >> > + *
> >> > + * For each VCPU this needs to be atomic, as we need *all* newly
> >> > + * enabled IRQs in be in the ap_list to determine the highest
> >> > + * priority one.
> >> > + * So grab the ap_list_lock, then iterate over all private IRQs
> >> and
> >> > + * all SPIs. Once the ap_list is updated, kick that VCPU to
> >> > + * forward any new IRQs to the guest.
> >> > */
> >> > + kvm_for_each_vcpu(cpuid, vcpu, kvm) {
> >> > + unsigned long flags;
> >> > + int i;
> >> >
> >> > - /* For now just kick all VCPUs, as the old code did. */
> >> > - vgic_kick_vcpus(kvm);
> >> > + raw_spin_lock_irqsave(&vcpu->arch.vgic_cpu.ap_list_lock,
> >> flags);
> >> > +
> >> > + for (i = 0; i < VGIC_NR_PRIVATE_IRQS; i++)
> >> > + queue_enabled_irq(kvm, vcpu, i);
> >> > +
> >> > + for (i = VGIC_NR_PRIVATE_IRQS;
> >> > + i < kvm->arch.vgic.nr_spis + VGIC_NR_PRIVATE_IRQS; i++)
> >> > + queue_enabled_irq(kvm, vcpu, i);
> >>
> >> On top of my questions above, what happens to LPIs?
> >
> > Oh dear. Looks like wishful thinking on my side ;-) Iterating over
> > all interrupts is probably not a good idea anymore.
> > Do you think this idea of having a list with group-disabled IRQs is a
> > better approach: In vgic_queue_irq_unlock, if a pending IRQ's group
> > is
> > enabled, it goes into the ap_list, if not, it goes into another list
> > instead. Then we would only need to consult this other list when a
> > group gets enabled. Both lists protected by the same ap_list_lock.
> > Does that make sense?
>
> I think that could work. One queue for each group, holding pending,
> enabled, group-disabled interrupts. Pending, disabled interrupts are
> not queued anywhere, just like today.
>
> The only snag is per-cpu interrupts. On which queue do they live?
> Do you have per-CPU queues? or a global one?
Yes, the idea was to have a per-VCPU "grp_dis_list" in addition to the ap_list, reusing the ap_list list_head in struct vgic_irq. vgic_queue_irq_unlock() would put them into *one* of those two lists, depending on their group-enabled status. When a group gets enabled, we just have to transfer the IRQs from grp_dis_list to ap_list.
But fleshing this out I was wondering if it couldn't be much simpler:
We ignore the group-enabled status most of the time, except in vgic_flush_lr_state(). So group-disabled IRQs *would go* to the ap_list (when they are otherwise pending|active and enabled), but would be skipped when eventually populating the LRs. vgic_prune_ap_list would also not touch them, so they would stay in the ap_list (unless removed for other reasons).
That might raise some eyebrows (because we keep IRQs in the ap_list which are not ready), but would require only minimal changes and avoid all kind of nasty/racy code to be added. The only downside I see is that the ap_list could potentially be much longer, but we could change the sorting algorithm if needed to keep group-disabled IRQs at the end, at which point it wouldn't really matter.
Do you see any problem with that approach? Alex seemed to remember that you had an objection against a very similar (if not identical) idea before.
> >> And if a group has
> >> been disabled, how do you retire these interrupts from the AP list?
> >
> > This is done above: we kick the respective VCPU and rely on
> > vgic_prune_ap_list() to remove them (that uses vgic_target_oracle(),
> > which in turn checks vgic_irq_is_grp_enabled()).
>
> But what if the CPU isn't running? Kicking it isn't going to do much,
> is it?
Not directly, but in either approach that would be handled similar to disabled interrupts: once the VCPU runs, they would *not* end up in LRs (because we check the oracle before), and would be cleaned up in prune() once the guest exits (at least for the original approach).
Cheers,
Andre.
