On Tue, Sep 06, 2016 at 09:28:40AM +0100, Marc Zyngier wrote:

In a number of cases, KVM cannot give access direct access to the
GICv2 GICV region, either because GICV is not page aligned, or its
size is not a multiple of the page size. This is especially visible
with 16kB/64kB pages and the original GIC-400 layout where each region
is only 4k aligned.

Instead of disabling KVM altogether (which is the current behaviour),
there is some value in trapping each guest GICV access, performing the
access as quickly as possible at EL2, and resuming the guest. This
allows us to keep KVM enabled on this HW.

Implementation wise, this is done with a static key controlling the
workaround being enabled, hence coming at zero cost (well, an extra
nop on the exit hot path) for unaffected platforms. On the affected
HW, I've measured a 10 to 15% overhead for a self-IPI test, which is
pretty bad, but still much better than not having a GIC at all.

There is two pending issues:

- A failed write to GICV ends up being forwarded to userspace. This
  will be addressed in a follow-up series where we deal with injecting
  vSError in the guest

- Skipping instructions (as we do when emulating anything) breaks
  things like guest single-step and watchpoints. This is a long
  standing problem, and someone should probably have a look at
  it. Alex?

Tested on Juno-r1 with 64kB pages.

I also tested this on TC2 to ensure we didn't regress the 32-bit side.

Applied the series.
-Christoffer