[RFC v3 00/10] Provide the EL1 physical timer to the VM

[Christoffer Dall]

On Thu, Feb 02, 2017 at 09:51:13AM -0500, Jintack Lim wrote:
> On Thu, Feb 2, 2017 at 7:31 AM, Christoffer Dall <cdall at linaro.org> wrote:
> > Hi Jintack,
> >
> > On Wed, Feb 01, 2017 at 12:43:00PM -0500, Jintack Lim wrote:
> >> The ARM architecture defines the EL1 physical timer and the virtual timer,
> >> and it is reasonable for an OS to expect to be able to access both.
> >> However, the current KVM implementation does not provide the EL1 physical
> >> timer to VMs but terminates VMs on access to the timer.
> >>
> >> This patch series enables VMs to use the EL1 physical timer through
> >> trap-and-emulate.  The KVM host emulates each EL1 physical timer register
> >> access and sets up the background timer accordingly.  When the background
> >> timer expires, the KVM host injects EL1 physical timer interrupts to the
> >> VM.  Alternatively, it's also possible to allow VMs to access the EL1
> >> physical timer without trapping.  However, this requires somehow using the
> >> EL2 physical timer for the Linux host while running the VM instead of the
> >> EL1 physical timer.  Right now I just implemented trap-and-emulate because
> >> this was straightforward to do, and I leave it to future work to determine
> >> if transferring the EL1 physical timer state to the EL2 timer provides any
> >> performance benefit.
> >>
> >> This feature will be useful for any OS that wishes to access the EL1
> >> physical timer. Nested virtualization is one of those use cases. A nested
> >> hypervisor running inside a VM would think it has full access to the
> >> hardware and naturally tries to use the EL1 physical timer as Linux would
> >> do. Other nested hypervisors may try to use the EL2 physical timer as Xen
> >> would do, but supporting the EL2 physical timer to the VM is out of scope
> >> of this patch series. This patch series will make it easy to add the EL2
> >> timer support in the future, though.
> >>
> >> Note that Linux VMs booting in EL1 will be unaffected by this patch series
> >> and will continue to use only the virtual timer and this patch series will
> >> therefore not introduce any performance degredation as a result of
> >> trap-and-emulate.
> >>
> >> v2 => v3:
> >>  - Rebase on kvmarm/queue
> >>  - Take kvm->lock to synchronize cntvoff across all vtimers
> >>  - Remove unnecessary function parameters
> >>  - Add comments
> >
> > I just gave v3 a test run on my TC2 (32-bit platform) and my guest
> > quickly locks up trying to run cyclictest or when booting the machine it
> > stalls with RCU timeouts.
> 
> Ok. It's my fault not to specify that the emulated physical timer is
> supported/tested on arm64.
> On 32-bit platform, it is supposed to show the same behavior as
> before, but I haven't tested.
> Were you using the physical timer or the virtual timer for the guest?
> 
> >
> > Could you have a look?
> 
> Sure, I'll have a look. I don't have access to my Cubietruck today,
> but I can work on that tomorrow.
> 

Don't bother, I've figured this out for you.

You need the following fixup to your patch:

diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c
index 93c811c..35d7100 100644
--- a/virt/kvm/arm/arch_timer.c
+++ b/virt/kvm/arm/arch_timer.c
@@ -410,14 +410,21 @@ int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
 }
 
 /* Make the updates of cntvoff for all vtimer contexts atomic */
-static void update_vtimer_cntvoff(struct kvm *kvm, u64 cntvoff)
+static void update_vtimer_cntvoff(struct kvm_vcpu *vcpu, u64 cntvoff)
 {
 	int i;
-	struct kvm_vcpu *vcpu;
+	struct kvm *kvm = vcpu->kvm;
+	struct kvm_vcpu *tmp;
 
 	mutex_lock(&kvm->lock);
-	kvm_for_each_vcpu(i, vcpu, kvm)
-		vcpu_vtimer(vcpu)->cntvoff = cntvoff;
+	kvm_for_each_vcpu(i, tmp, kvm)
+		vcpu_vtimer(tmp)->cntvoff = cntvoff;
+
+	/*
+	 * When called from the vcpu create path, the CPU being created is not
+	 * included in the loop above, so we just set it here as well.
+	 */
+	vcpu_vtimer(vcpu)->cntvoff = cntvoff;
 	mutex_unlock(&kvm->lock);
 }
 
@@ -426,7 +433,7 @@ void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
 	struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
 
 	/* Synchronize cntvoff across all vtimers of a VM. */
-	update_vtimer_cntvoff(vcpu->kvm, kvm_phys_timer_read());
+	update_vtimer_cntvoff(vcpu, kvm_phys_timer_read());
 	vcpu_ptimer(vcpu)->cntvoff = 0;
 
 	INIT_WORK(&timer->expired, kvm_timer_inject_irq_work);
@@ -448,7 +455,7 @@ int kvm_arm_timer_set_reg(struct kvm_vcpu *vcpu, u64 regid, u64 value)
 		vtimer->cnt_ctl = value;
 		break;
 	case KVM_REG_ARM_TIMER_CNT:
-		update_vtimer_cntvoff(vcpu->kvm, kvm_phys_timer_read() - value);
+		update_vtimer_cntvoff(vcpu, kvm_phys_timer_read() - value);
 		break;
 	case KVM_REG_ARM_TIMER_CVAL:
 		vtimer->cnt_cval = value;

This is an amuzing one.

Thanks,
-Christoffer