[RFC v2 05/10] KVM: arm/arm64: Initialize the emulated EL1 physical timer

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

On Sun, Jan 29, 2017 at 12:07:48PM +0000, Marc Zyngier wrote:
> On Fri, Jan 27 2017 at 01:04:55 AM, Jintack Lim <jintack@cs.columbia.edu> wrote:
> > Initialize the emulated EL1 physical timer with the default irq number.
> >
> > Signed-off-by: Jintack Lim <jintack@cs.columbia.edu>
> > ---
> >  arch/arm/kvm/reset.c         | 9 ++++++++-
> >  arch/arm64/kvm/reset.c       | 9 ++++++++-
> >  include/kvm/arm_arch_timer.h | 3 ++-
> >  virt/kvm/arm/arch_timer.c    | 9 +++++++--
> >  4 files changed, 25 insertions(+), 5 deletions(-)
> >
> > diff --git a/arch/arm/kvm/reset.c b/arch/arm/kvm/reset.c
> > index 4b5e802..1da8b2d 100644
> > --- a/arch/arm/kvm/reset.c
> > +++ b/arch/arm/kvm/reset.c
> > @@ -37,6 +37,11 @@
> >  	.usr_regs.ARM_cpsr = SVC_MODE | PSR_A_BIT | PSR_I_BIT | PSR_F_BIT,
> >  };
> >  
> > +static const struct kvm_irq_level cortexa_ptimer_irq = {
> > +	{ .irq = 30 },
> > +	.level = 1,
> > +};
> 
> At some point, we'll have to make that discoverable/configurable. Maybe
> the time for a discoverable arch timer has come (see below).
> 
> > +
> >  static const struct kvm_irq_level cortexa_vtimer_irq = {
> >  	{ .irq = 27 },
> >  	.level = 1,
> > @@ -58,6 +63,7 @@ int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> >  {
> >  	struct kvm_regs *reset_regs;
> >  	const struct kvm_irq_level *cpu_vtimer_irq;
> > +	const struct kvm_irq_level *cpu_ptimer_irq;
> >  
> >  	switch (vcpu->arch.target) {
> >  	case KVM_ARM_TARGET_CORTEX_A7:
> > @@ -65,6 +71,7 @@ int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> >  		reset_regs = &cortexa_regs_reset;
> >  		vcpu->arch.midr = read_cpuid_id();
> >  		cpu_vtimer_irq = &cortexa_vtimer_irq;
> > +		cpu_ptimer_irq = &cortexa_ptimer_irq;
> >  		break;
> >  	default:
> >  		return -ENODEV;
> > @@ -77,5 +84,5 @@ int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> >  	kvm_reset_coprocs(vcpu);
> >  
> >  	/* Reset arch_timer context */
> > -	return kvm_timer_vcpu_reset(vcpu, cpu_vtimer_irq);
> > +	return kvm_timer_vcpu_reset(vcpu, cpu_vtimer_irq, cpu_ptimer_irq);
> >  }
> > diff --git a/arch/arm64/kvm/reset.c b/arch/arm64/kvm/reset.c
> > index e95d4f6..d9e9697 100644
> > --- a/arch/arm64/kvm/reset.c
> > +++ b/arch/arm64/kvm/reset.c
> > @@ -46,6 +46,11 @@
> >  			COMPAT_PSR_I_BIT | COMPAT_PSR_F_BIT),
> >  };
> >  
> > +static const struct kvm_irq_level default_ptimer_irq = {
> > +	.irq	= 30,
> > +	.level	= 1,
> > +};
> > +
> >  static const struct kvm_irq_level default_vtimer_irq = {
> >  	.irq	= 27,
> >  	.level	= 1,
> > @@ -104,6 +109,7 @@ int kvm_arch_dev_ioctl_check_extension(struct kvm *kvm, long ext)
> >  int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> >  {
> >  	const struct kvm_irq_level *cpu_vtimer_irq;
> > +	const struct kvm_irq_level *cpu_ptimer_irq;
> >  	const struct kvm_regs *cpu_reset;
> >  
> >  	switch (vcpu->arch.target) {
> > @@ -117,6 +123,7 @@ int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> >  		}
> >  
> >  		cpu_vtimer_irq = &default_vtimer_irq;
> > +		cpu_ptimer_irq = &default_ptimer_irq;
> >  		break;
> >  	}
> >  
> > @@ -130,5 +137,5 @@ int kvm_reset_vcpu(struct kvm_vcpu *vcpu)
> >  	kvm_pmu_vcpu_reset(vcpu);
> >  
> >  	/* Reset timer */
> > -	return kvm_timer_vcpu_reset(vcpu, cpu_vtimer_irq);
> > +	return kvm_timer_vcpu_reset(vcpu, cpu_vtimer_irq, cpu_ptimer_irq);
> >  }
> > diff --git a/include/kvm/arm_arch_timer.h b/include/kvm/arm_arch_timer.h
> > index 69f648b..a364593 100644
> > --- a/include/kvm/arm_arch_timer.h
> > +++ b/include/kvm/arm_arch_timer.h
> > @@ -59,7 +59,8 @@ struct arch_timer_cpu {
> >  int kvm_timer_enable(struct kvm_vcpu *vcpu);
> >  void kvm_timer_init(struct kvm *kvm);
> >  int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
> > -			 const struct kvm_irq_level *irq);
> > +			 const struct kvm_irq_level *virt_irq,
> > +			 const struct kvm_irq_level *phys_irq);
> >  void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu);
> >  void kvm_timer_flush_hwstate(struct kvm_vcpu *vcpu);
> >  void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu);
> > diff --git a/virt/kvm/arm/arch_timer.c b/virt/kvm/arm/arch_timer.c
> > index f72005a..0f6e935 100644
> > --- a/virt/kvm/arm/arch_timer.c
> > +++ b/virt/kvm/arm/arch_timer.c
> > @@ -329,9 +329,11 @@ void kvm_timer_sync_hwstate(struct kvm_vcpu *vcpu)
> >  }
> >  
> >  int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
> > -			 const struct kvm_irq_level *irq)
> > +			 const struct kvm_irq_level *virt_irq,
> > +			 const struct kvm_irq_level *phys_irq)
> >  {
> >  	struct arch_timer_context *vtimer = vcpu_vtimer(vcpu);
> > +	struct arch_timer_context *ptimer = vcpu_ptimer(vcpu);
> >  
> >  	/*
> >  	 * The vcpu timer irq number cannot be determined in
> > @@ -339,7 +341,8 @@ int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
> >  	 * kvm_vcpu_set_target(). To handle this, we determine
> >  	 * vcpu timer irq number when the vcpu is reset.
> >  	 */
> > -	vtimer->irq.irq = irq->irq;
> > +	vtimer->irq.irq = virt_irq->irq;
> > +	ptimer->irq.irq = phys_irq->irq;
> >  
> >  	/*
> >  	 * The bits in CNTV_CTL are architecturally reset to UNKNOWN for ARMv8
> > @@ -348,6 +351,7 @@ int kvm_timer_vcpu_reset(struct kvm_vcpu *vcpu,
> >  	 * the ARMv7 architecture.
> >  	 */
> >  	vtimer->cnt_ctl = 0;
> > +	ptimer->cnt_ctl = 0;
> >  	kvm_timer_update_state(vcpu);
> >  
> >  	return 0;
> > @@ -369,6 +373,7 @@ void kvm_timer_vcpu_init(struct kvm_vcpu *vcpu)
> >  	struct arch_timer_cpu *timer = &vcpu->arch.timer_cpu;
> >  
> >  	update_vtimer_cntvoff(vcpu, kvm_phys_timer_read());
> > +	vcpu_ptimer(vcpu)->cntvoff = 0;
> 
> This is quite contentious, IMHO. Do we really want to expose the delta
> between the virtual and physical counters? That's a clear indication to
> the guest that it is virtualized. I"m not sure it matters, but I think
> we should at least make a conscious choice, and maybe give the
> opportunity to userspace to select the desired behaviour.
> 

So my understanding of the architecture is that you should always have
two timer/counter pairs available at EL1.  They may be synchronized, and
they may not.  If you want an accurate reading of wall clock time, look
at the physical counter, and that can generally be expected to be fast,
precise, and syncrhonized (on working hardware, of course).

Now, there can be a constant or potentially monotonically increasing
offset between the physial and virtual counters, which may mean you're
running under a hypervisor or (in the first case) that firmware
programmed or neglected to program cntvoff.  I don't think it's an
inherent problem to expose that difference to a guest, and I think it's
more important that reading the physical counter is fast and doesn't
trap.

The question is which contract we can have with a guest OS, and which
legacy we have to keep supporting (Linux, UEFI, ?).

Probably Linux should keep relying on the virtual counter/timer only,
unless something is advertised in DT/ACPI, about it being able to use
the physical timer/counter pair, even when booted at EL1.  We could
explore the opportunities to build on that to let the guest figure
out stolen time by reading the two counters and by programming the
proper timer depending on the desired semantics (i.e. virtual or
physical time).

In terms of this patch, I actually think it's fine, but we may need to
build something more on top later.  It is possible, though, that I'm
entirely missing the point about Linux timekeeping infrastructure and
that my reading of the architecture is bogus.

What do you think?

Thanks,
-Christoffer