Hi Simon,

On 29.04.2016 12:35, Marc Zyngier wrote:

On Fri, 29 Apr 2016 09:43:45 +1000
Simon Horman [off-list ref] wrote:

quoted

[Cc Mark Zyngier, linux-arm-kernel]

Hi Dirk,

On Thu, Apr 28, 2016 at 07:41:57AM +0200, Dirk Behme wrote:

quoted

Hi Simon,

On 28.04.2016 01:30, Simon Horman wrote:

quoted

Hi Dirk,

I understand that there is an issue here but I'm not yet able
to convince myself that this is the correct solution.

In revision r0p1 of the CoreLink GIC-400 Generic Interrupt Controller
Technical Reference Manual[1] I see in Section 3.2. "GIC-400 register map"
that the size of both the CPU interfaces and Virtual CPU interfaces are
0x2000 bytes. And assuming that the hardware follows the specification it
appears that DT is correctly describing the hardware.

I think you are missing the details described by ARM in

http://xenbits.xen.org/gitweb/?p=xen.git;a=commit;h=21550029f709072aacf3b9


Maybe Julien could help if you have some more doubts?

I guess I am confused.

I see that there is now handling of the case where the region size is
128Kbytes. But I'm still not seeing the bit which describes that the
GIC-400  has a region size of 128Kbytes. Perhaps the later is somehow
implied by the former. Or perhaps I need to check with the hw team.

Please have a look at the SBSA document, and in particular the
Appendix-F (registration and selling your soul required - only kidding):

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.den0029/index.html

This requires that, in order for the two halves of GICV to be trappable
*separately* by a hypervisor using 64kB pages at Stage-2, the two 4kB
pages that describe that region are aliased as such:
- the first 4kB page is aliased 16 times over a 64kB region
- the second 4kB page is aliased 16 times over another contiguous 64kB
   region

This means that your GIC is indeed covering a 128kB region, with the
mapping corresponding to the GICv2 memory map located at offset 0xf000
from the base of that 128kB region. Also, this GICV requirement also
applies to GICC (most likely because the two regions use the same
decoding logic).

The OS must of course be aware of this (see gic_check_eoimode in the
GIC driver). Of course, almost nobody got that right (I only know of
the APM Xgene-1 so far). If you actually did, great!

Also, the ACPI spec fails to recognize this by not providing the length
of the region, meaning that those who got it right with DT are likely
to get it wrong with ACPI, and vice-versa.

It's a wonderful world.

Could this patch be applied, then?

Best regards

Dirk