On Wed, Nov 30, 2016 at 07:10:51AM +0000, Peter Maydell wrote:
> On 29 November 2016 at 21:09, Christoffer Dall
> <christoffer.d...@linaro.org> wrote:
> > Actually, I'm not sure what the semantics of the line level ioctl should
> > be for edge-triggered interrupts?  My inclination is that it shouldn't
> > have any effect at this point, but that would mean that at this point we
> > should only set the pending variable and try to queue the interrupt if
> > the config is level.  But that also means that when we set the config
> > later, we need to try to queue the interrupt, which we don't do
> > currently, because we rely on the guest not fiddling with the config of
> > an enabled interrupt.
> >
> > Could it be considered an error if user space tries to set the level for
> > an edge-triggered interrupt and therefore something we can just ignore
> > and assume that the corresponing interrupt will be configured as a
> > level-triggered one later?
> Userspace will always read the line-level values out and write
> them back for migration, and I'd rather not make it have to
> do cross-checks against whether the interrupt is edge or level
> triggered to see whether it should write the level values into
> the kernel. Telling the kernel the level for an edge-triggered
> interrupt should be a no-op because it doesn't have any effect
> on pending status.
> > In any case we probably need to clarify the ABI in terms of this
> > particular KVM_DEV_AR_VGIC_GRP_LEVEL_INFO group and how it relates to
> > the config of edge vs. level of interrupts and ordering on restore...
> IIRC the QEMU code restores the config first. (There's a similar
> ordering thing for GICv2 where we have to restore GICD_ICFGRn before

So it sounds to me that we should add a note in the Documentation like

diff --git a/Documentation/virtual/kvm/devices/arm-vgic-v3.txt 
index 9348b3c..7bac20a 100644
--- a/Documentation/virtual/kvm/devices/arm-vgic-v3.txt
+++ b/Documentation/virtual/kvm/devices/arm-vgic-v3.txt
@@ -193,6 +193,11 @@ Groups:
        Bit[n] indicates the status for interrupt vINTID + n.
+       Getting or setting the level info for an edge-triggered interrupt is
+       not guaranteed to work.  To restore the complete state of the VGIC, the
+       configuration (edge/level) of interrupts must be restored before
+       restoring the level.
     SGIs and any interrupt with a higher ID than the number of interrupts
     supported, will be RAZ/WI.  LPIs are always edge-triggered and are
     therefore not supported by this interface.

Vijay, this means that the first block in your if-statement should only
set pending and queue the interrupt if the interrupt is a level
triggered one.

(Peter, I thought you once argued that it was important for user space
to be able to save/restore the state without any ordering requirements,
but I may have misunderstood.  It is still the option to add something
like the above to the docs but also do our best to allow any ordering of
level/config, but it becomes slightly more invasive.)

kvmarm mailing list

Reply via email to