Re: [Qemu-devel] [PATCH v7 6/6] Add gicversion option to virt machine

[Peter Maydell]

On 3 August 2015 at 12:41, Shlomo Pongratz <address@hidden> wrote:
>
>
> On Monday, August 3, 2015, Peter Maydell <address@hidden> wrote:
>> I'm surprised we tell the CPU about the GIC pointer for the
>> system register stuff -- I was expecting that we'd give the
>> GIC a CPU pointer. (We could in theory implement some
>> equivalent of the architected protocol between the redistributors
>> and the CPU interfaces, but I think that would be overkill.)

> The problem is that each function added as a system's instruction helper to
> target-arm/cpu64.c has the signature of void set(CPUARMState *env,
> ARMCPRegInfo *ri, uint64_t value) or uint64_t get(CPUARMState *env,
> ARMCPRegInfo *ri)
> I just mimicked the way armv7m_nvic_XXXX API works.

The v7M NVIC is a terrible example to copy -- it is ancient
code mostly written before QEMU acquired various useful
abstraction layers, and has received very little maintenance
since then.

> So in a sense the CPU must be familiar with the GIC (as an opaque object of
> course).

If the GIC just registers a set of system register information with
the CPU then the CPU doesn't need to know that it's the GIC
providing those system register functions. (You can use
define_arm_cp_regs_with_opaque() to define registers such
that your callbacks can fish an opaque data pointer back out
of ri->opaque, so you can get back to the GIC data structures.)

-- PMM