Re: [RFC 2/2] ARM: KVM: Enable in-kernel timers with user space gic

[Christoffer Dall <christoffer.dall@linaro.org>]

On Wed, Nov 02, 2016 at 04:40:35PM +0100, Alexander Graf wrote:
> On 11/01/2016 12:35 PM, Peter Maydell wrote:
> >On 29 October 2016 at 22:10, Alexander Graf <agraf@suse.de> wrote:

[...]

> >
> >>+        cpu->timer_irq_level = run->s.regs.timer_irq_level;
> >>+    }
> >>+
> >>      return MEMTXATTRS_UNSPECIFIED;
> >>  }
> >Does this code do the right thing across a vcpu reset or
> >a full-system reset?
> 
> Good question. I'm not 100% sure - but I don't know for sure whether
> it's guaranteed without user space irqchip even.
> 
> In essence, the code above merely synchronizes kvm state to qemu
> state and is fully unaffected from any reset sequence. This is good,
> as the line status is transient. So from a QEMU pov, we really only
> copy the state of the vcpu interrupt line into the QEMU interrupt
> line. Pulling that line down would be responsibility of the
> KVM_ARM_VCPU_INIT ioctl if it also clears the timer registers I
> guess.
> 
> However, I don't see any clearing of cntv_ctrl inside KVM or from
> QEMU. How do we ensure that the irq active bit is off on reset?


In kvm_timer_vcpu_reset we cset cntv_ctl = 0, and that function gets
called from the PSCI handler or whenever userspace calls the set target
ioctl thingy.

> 
> The other part that could get in the way of working system reset is
> the interrupt controller emulation itself which resets all internal
> irq line state. So on reset we'd always end up with the irq line
> down from a gic pov, but with the vtimer line pending or not pending
> depending on previous state. I doubt it's really going to hurt
> though.

I suppose it should resample the line, but if the GIC clears everything
and the arch timer line goes down, you're in the right starting state
again.  Right?

-Christoffer