ARM GIC DT binding reg block mismatch? (Re: [PATCH v11 1/8] arm64: renesas: r8a7795: Add Renesas R8A7795 SoC support)

[Marc Zyngier]

On 15/02/16 10:35, Geert Uytterhoeven wrote:
> Hi Marc,
> 
> On Mon, Feb 15, 2016 at 9:45 AM, Marc Zyngier <marc.zyngier at arm.com> wrote:
>> On 15/02/16 08:16, Geert Uytterhoeven wrote:
>>> On Wed, Dec 9, 2015 at 9:23 AM, Geert Uytterhoeven <geert at linux-m68k.org> wrote:
>>>> On Tue, Nov 3, 2015 at 3:28 PM, Mark Rutland <mark.rutland at arm.com> wrote:
>>>>> On Wed, Oct 21, 2015 at 03:34:39PM +0200, Geert Uytterhoeven wrote:
>>>>>> On Thu, Oct 15, 2015 at 12:58 PM, Mark Rutland <mark.rutland at arm.com> wrote:
>>>>>>>>> +           gic: interrupt-controller at 0xf1010000 {
>>>>>>>> +                     compatible = "arm,gic-400";
>>>>>>>> +                     #interrupt-cells = <3>;
>>>>>>>> +                     #address-cells = <0>;
>>>>>>>> +                     interrupt-controller;
>>>>>>>> +                     reg = <0x0 0xf1010000 0 0x1000>,
>>>>>>>> +                           <0x0 0xf1020000 0 0x2000>;
>>>>>>>> +                     interrupts = <GIC_PPI 9
>>>>>>>> +                                     (GIC_CPU_MASK_SIMPLE(1) | IRQ_TYPE_LEVEL_HIGH)>;
>>>>>>>> +             };
>>>>>>>
>>>>>>> No GICH and GICV?
>>>>>>
>>>>>> These seem to be defined in the "arm,gic-v3" DT bindings only, while this is
>>>>>> an "arm,gic-400" (GICD_IIDR 0x0200043b)?
>>>>>
>>>>> See the "GIC virtualization extensions (VGIC)" section in
>>>>> Documentation/devicetree/bindings/arm/gic.txt
>>>>
>>>> DDI0471B_gic400_r0p1_trm.pdf says:
>>>>
>>>>     Address range GIC-400 functional block
>>>>     A. 0x0000 - 0x0FFF Reserved
>>>>     B. 0x1000 - 0x1FFF Distributor
>>>>     C. 0x2000 - 0x3FFF CPU interfaces
>>>>     D. 0x4000 - 0x4FFF Virtual interface control block, for the processor that
>>>>                        is performing the access
>>>>     E. 0x5000 - 0x5FFF Virtual interface control block, for the processor
>>>>                        selected by address bits [11:9]
>>>>     F. 0x6000 - 0x7FFF Virtual CPU interfaces
>>>>
>>>> The DT binding document says:
>>>>   1. The  first region is the GIC distributor register base and size.
>>>>   2. The 2nd region is the GIC cpu interface register base and size.
>>>>   3. The first additional region is the GIC virtual interface control register
>>>>      base and size.
>>>>   4. The 2nd additional region is the GIC virtual cpu interface register base
>>>>      and size.
>>>>
>>>> Matching with the example:
>>>>
>>>>         interrupt-controller at 2c001000 {
>>>>                 compatible = "arm,cortex-a15-gic";
>>>>                 #interrupt-cells = <3>;
>>>>                 interrupt-controller;
>>>>                 reg = <0x2c001000 0x1000>,
>>>>                       <0x2c002000 0x1000>,
>>>>                       <0x2c004000 0x2000>,
>>>>                       <0x2c006000 0x2000>;
>>>>                 interrupts = <1 9 0xf04>;
>>>>         };
>>>>
>>>> This means:
>>>>   - reg entry 1. covers address range B,
>>>>   - reg entry 2. covers address range C,
>>>>   - reg entry 3. covers address ranges D _and_ E,
>>>>   - reg entry 4. covers address range F.
>>>>
>>>> On R-Car Gen3, the base addresses are:
>>>>
>>>>     Distributor             : 0xF101_0000
>>>>     CPU interfaces          : 0xF102_0000
>>>>     Virtual interfaces      : 0xF104_0000
>>>>     Virtual interfaces      : 0xF105_0000
>>>>     Virtual CPU interfaces  : 0xF106_0000
>>>>
>>>> Note the additional multiplication factor of 16 in the offsets relative to
>>>> the base address 0xf1000000 (e.g. 0x50000 instead of 0x5000).
>>>>
>>>> As address ranges D and E are merged in a single reg entry, how is the GIC
>>>> driver supposed to know about this multiplication factor?
>>
>> The answer is very simple, the GIC driver doesn't give a damn about the
>> second part of the GICH region, because it is absolutely unusable for
>> any realistic use-case. Only the banked version of GICH is of any
>> relevance (the first 512 bytes, in essence).
>>
>> Aligning the GIC regions on 64kB boundaries is documented in the SBSA
>> specification, independently of the GIC400 documentation.
> 
> If I understand the SBSA spec correctly (BTW, arm,gic.txt doesn't use the
> "GICC" terminology, unlike arm,gic-v3.txt), that means reg entry 3 should be
> "<0xf104f000 0x2000>", so it covers the aliased last 4 KiB of address range D,
> and the first 4 KiB of address range E. I.e.
> 
>                         reg = <0x0 0xf1010000 0 0x1000>,
>                              <0x0 0xf1020000 0 0x2000>,
>                              <0x0 0xf104f000 0 0x2000>,
>                              <0x0 0xf1060000 0 0x2000>;
> 
> Is that correct?

My preference would be to expose the full 128kB of the region - assuming
your HW actually supports the SBSA aliasing.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...