On Tue, 02 Apr 2019 09:22:29 +0100,
Auger Eric <eric.auger@xxxxxxxxxx> wrote:
Hi Marc,
On 4/2/19 9:24 AM, Marc Zyngier wrote:
When disabling LPIs (for example on reset) at the redistributor
level, it is expected that LPIs that was pending in the CPU
interface are eventually retired.
Currently, this is not what is happening, and these LPIs will
stay in the ap_list, eventually being acknowledged by the vcpu
(which didn't quite expect this behaviour).
The fix is thus to retire these LPIs from the list of pending
interrupts as we disable LPIs.
Reported-by: Heyi Guo <guoheyi@xxxxxxxxxx>
Fixes: 0e4e82f154e3 ("KVM: arm64: vgic-its: Enable ITS emulation as a virtual MSI controller")
Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
---
Heyi,
Can you please give this alternative patch a go on your setup?
This should be a much better fit than my original hack.
Thanks,
M.
virt/kvm/arm/vgic/vgic-mmio-v3.c | 3 +++
virt/kvm/arm/vgic/vgic.c | 24 ++++++++++++++++++++++++
virt/kvm/arm/vgic/vgic.h | 1 +
3 files changed, 28 insertions(+)
diff --git a/virt/kvm/arm/vgic/vgic-mmio-v3.c b/virt/kvm/arm/vgic/vgic-mmio-v3.c
index 4a12322bf7df..9f4843fe9cda 100644
--- a/virt/kvm/arm/vgic/vgic-mmio-v3.c
+++ b/virt/kvm/arm/vgic/vgic-mmio-v3.c
@@ -200,6 +200,9 @@ static void vgic_mmio_write_v3r_ctlr(struct kvm_vcpu *vcpu,
vgic_cpu->lpis_enabled = val & GICR_CTLR_ENABLE_LPIS;
+ if (was_enabled && !vgic_cpu->lpis_enabled)
+ vgic_flush_pending_lpis(vcpu);
+
if (!was_enabled && vgic_cpu->lpis_enabled)
vgic_enable_lpis(vcpu);
}
diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
index 3af69f2a3866..3f429cf2f694 100644
--- a/virt/kvm/arm/vgic/vgic.c
+++ b/virt/kvm/arm/vgic/vgic.c
@@ -151,6 +151,30 @@ void vgic_put_irq(struct kvm *kvm, struct vgic_irq *irq)
kfree(irq);
}
+void vgic_flush_pending_lpis(struct kvm_vcpu *vcpu)
+{
+ struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
+ struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
+ struct vgic_irq *irq, *tmp;
+ unsigned long flags;
+
+ raw_spin_lock_irqsave(&dist->lpi_list_lock, flags);
Do you need to hold the lpi_list_lock here as it is taken in vgic_put_irq.
You're right, this is absolutely silly, as it violates the rules at
the top of this very file. Don't write that kind of patches while
being severely jet-lagged. It should be better with this on top.
diff --git a/virt/kvm/arm/vgic/vgic.c b/virt/kvm/arm/vgic/vgic.c
index 3f429cf2f694..191deccf60bf 100644
--- a/virt/kvm/arm/vgic/vgic.c
+++ b/virt/kvm/arm/vgic/vgic.c
@@ -153,13 +153,11 @@ void vgic_put_irq(struct kvm *kvm, struct vgic_irq *irq)
void vgic_flush_pending_lpis(struct kvm_vcpu *vcpu)
{
- struct vgic_dist *dist = &vcpu->kvm->arch.vgic;
struct vgic_cpu *vgic_cpu = &vcpu->arch.vgic_cpu;
struct vgic_irq *irq, *tmp;
unsigned long flags;
- raw_spin_lock_irqsave(&dist->lpi_list_lock, flags);
- raw_spin_lock(&vgic_cpu->ap_list_lock);
+ raw_spin_lock_irqsave(&vgic_cpu->ap_list_lock, flags);
list_for_each_entry_safe(irq, tmp, &vgic_cpu->ap_list_head, ap_list) {
if (irq->intid >= VGIC_MIN_LPI) {
@@ -171,8 +169,7 @@ void vgic_flush_pending_lpis(struct kvm_vcpu *vcpu)
}
}
- raw_spin_unlock(&vgic_cpu->ap_list_lock);
- raw_spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
+ raw_spin_unlock_irqrestore(&vgic_cpu->ap_list_lock, flags);
}
void vgic_irq_set_phys_pending(struct vgic_irq *irq, bool pending)
+ raw_spin_lock(&vgic_cpu->ap_list_lock);
+
+ list_for_each_entry_safe(irq, tmp, &vgic_cpu->ap_list_head, ap_list) {
+ if (irq->intid >= VGIC_MIN_LPI) {
+ raw_spin_lock(&irq->irq_lock);
+ list_del(&irq->ap_list);
+ irq->vcpu = NULL;
+ raw_spin_unlock(&irq->irq_lock);
+ vgic_put_irq(vcpu->kvm, irq);
+ }
+ }
+
+ raw_spin_unlock(&vgic_cpu->ap_list_lock);
+ raw_spin_unlock_irqrestore(&dist->lpi_list_lock, flags);
A concern I have is how does this live with
vgic_v3_save_pending_tables/its_sync_lpi_pending_table ctrl flow. I need
to trace things.
It shouldn't change a thing. The only thing this patch tries to do is
to make sure that LPIs are not kept on the ap_list when they have been
turned off at the redistributor level. The pending bits are still
stored, and save/sync should work as expected (famous last words...).
Thanks,
M.
