Re: [PATCH] target/arm: do not use cc->do_interrupt for KVM directly

[Philippe Mathieu-Daudé]

On 12/7/20 10:50 PM, Peter Maydell wrote:
> On Mon, 7 Dec 2020 at 21:26, Eduardo Habkost <ehabkost@redhat.com> wrote:
>> My understanding is that there's no reason for ARM KVM to use
>> another approach, and that CPUClass.do_interrupt is not really
>> TCG-specific.
>>
>> Do we have any case where the CPUClass.do_interrupt
>> implementation is really TCG-specific, or it is just a
>> coincidence that most other accelerators simply don't to call the
>> method?  It looks like the only cases where the
>> CPUClass.do_interrupt assignment is conditional on CONFIG_TCG are
>> i386 and s390x.
> 
> Looking at PPC, its kvm_handle_debug() function does a
> direct call to ppc_cpu_do_interrupt(). So the code of
> its do_interrupt method must be ok-for-KVM, it's just that
> it doesn't use the method pointer. (It's doing the same thing
> Arm is -- if a debug event turns out not to be for QEMU itself,
> inject a suitable exception into the guest.)
> 
> So of our 5 KVM-supporting architectures:
> 
>  * i386 and s390x have kernel APIs for "inject suitable
>    exception", don't need to call do_interrupt, and make
>    the cc->do_interrupt assignment only ifdef CONFIG_TCG,
>    so that the code for do_interrupt need not be compiled
>    into a KVM-only binary. (In both cases the code for the
>    function is in a source file that the meson.build puts
>    into the source list only if CONFIG_TCG)
>  * ppc and arm both need to use this code even in a KVM
>    only binary. Neither of them #ifdef the cc->do_interrupt
>    assignment, because there's not much point at the moment
>    if you're not going to try to compile out the code.
>    ppc happens to do a direct function call, and arm happens
>    to go via the cc->do_interrupt pointer, but I don't
>    think there's much significance in the choice either way.
>    In both cases, the only places making the call are within
>    architecture-specific KVM code.
>  * mips KVM does neither of these things, probably because it is
>    not sufficiently featureful to have run into the cases
>    where you might want to re-inject an exception and it's
>    not being sufficiently used in production for anybody to
>    have looked at minimising the amount of code in a
>    KVM-only QEMU binary for it.
> 
> So in conclusion we have a basically 50:50 split between
> "use the same do_interrupt code as TCG" and "have a kernel
> API to make the kernel do the work", plus one arch that
> probably hasn't had to make the choice yet.   ¯\_(ツ)_/¯

Why not introduce KVMCpuOperations similar to TCGCpuOperations
Claudio is introducing, and declare the do_interrupt(CPUState*)
in both structures?

Then we can assign the same handler to both fields, TCG keeps
calling cc->tcg->do_interrupt(), KVM calls cc->kvm->do_interrupt().
This allow building with a particular accelerator, while staying
compliant with the current 50:50 split...

> 
>> Oh, I thought you were arguing that CPUClass.do_interrupt is
>> not TCG_specific.
> 
> Well, I don't think it really is TCG-specific. But as
> a pragmatic thing, if these two lines in the Arm code
> are getting in the way of stopping us from having a
> useful compile-time check that code that's not supposed
> to call this method isn't calling it, I think the balance
> maybe leans towards just making the direct function call.
> I guess it depends whether you think people are likely to
> accidentally make cc->do_interrupt calls in non-target-specific
> code that gets used by KVM (which currently would crash if that
> code path is exercised on x86 or s390x, but under the
> proposed change would become a compile error).
> 
> thanks
> -- PMM
>