[PATCH 06/16] KVM: Add KVM_EXIT_SYSTEM_EVENT to user space API header

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

On Fri, Aug 29, 2014 at 06:39:09PM +0100, Peter Maydell wrote:
> On 25 May 2014 19:18, Christoffer Dall <christoffer.dall@linaro.org> wrote:
> > From: Anup Patel <anup.patel@linaro.org>
> >
> > Currently, we don't have an exit reason to notify user space about
> > a system-level event (for e.g. system reset or shutdown) triggered
> > by the VCPU. This patch adds exit reason KVM_EXIT_SYSTEM_EVENT for
> > this purpose. We can also inform user space about the 'type' and
> > architecture specific 'flags' of a system-level event using the
> > kvm_run structure.
> >
> > This newly added KVM_EXIT_SYSTEM_EVENT will be used by KVM ARM/ARM64
> > in-kernel PSCI v0.2 support to reset/shutdown VMs.
> 
> > --- a/Documentation/virtual/kvm/api.txt
> > +++ b/Documentation/virtual/kvm/api.txt
> > @@ -2740,6 +2740,21 @@ It gets triggered whenever both KVM_CAP_PPC_EPR are enabled and an
> >  external interrupt has just been delivered into the guest. User space
> >  should put the acknowledged interrupt vector into the 'epr' field.
> >
> > +               /* KVM_EXIT_SYSTEM_EVENT */
> > +               struct {
> > +#define KVM_SYSTEM_EVENT_SHUTDOWN       1
> > +#define KVM_SYSTEM_EVENT_RESET          2
> > +                       __u32 type;
> > +                       __u64 flags;
> > +               } system_event;
> > +
> > +If exit_reason is KVM_EXIT_SYSTEM_EVENT then the vcpu has triggered
> > +a system-level event using some architecture specific mechanism (hypercall
> > +or some special instruction). In case of ARM/ARM64, this is triggered using
> > +HVC instruction based PSCI call from the vcpu. The 'type' field describes
> > +the system-level event type. The 'flags' field describes architecture
> > +specific flags for the system-level event.
> 
> Talking with Ard I realised that there's actually a hole in the
> specification of this new ABI. Did we intend these shutdown
> and reset exits to be:
>  (1) requests from the guest for the shutdown/reset to be
>    scheduled in the near future (and we'll continue to execute
>    the guest until the shutdown actually happens)
>  (2) requests for shutdown/reset right now, with no further
>    guest instructions to be executed
> 
> ?
> 
> As currently implemented in QEMU we get behaviour (1),
> but I think the kernel PSCI implementation assumes
> behaviour (2). Who's right?
> 
For the arm/arm64 use of this API (currently the only one?) the host
would not break or anything like that if you keep executing the VM, but
the guest will expect that no other instructions are executed after this
call.

The PSCI spec states that it's the responsibility of the PSCI
implementation (here KVM), that "Implementation must ensure that all
cores are in a known state with caches cleaned".  I guess we don't need
to worry about the latter, but we could handle the former by pausing all
VCPUs prior to exiting with the SHUTDOWN system event.  In that
scenario, user space could choose to do either (1) or (2), but it gets a
little fishy with a reset if we set the pause flag, because we would
then at least need to specify in this ABI that this happens for
ARM/ARM64 on reset.

We could clarify this ABI to the fact that user space should not run any
VCPUs after receiving this event, but the above change should probably
be made anyhow, to make sure KVM implements PSCI as much as it can in
the kernel?

-Christoffer