Re: + stupid-hack-to-make-mainline-build.patch added to -mm tree

[Jeremy Fitzhardinge <jeremy@goop.org>]

Chris Wright wrote:
> * Jeremy Fitzhardinge (jeremy@goop.org) wrote:
>   
>> Chris Wright wrote:
>>     
>>> I agree with that, but I think that's esp. for things like create and launch
>>> new vcpu.  The IPI bit I'm not as clear on, nor running this all on native
>>> as well.
>>>   
>>>       
>> Well, native would fall back to using the existing arch/i386 versions of
>> those functions, so that's reasonably straightforward.
>>     
>
> It's the fact that we need to leave code in the kernel to run on native,
> but also do something dynamically with that same code when running
> paravirt that I'm referring to.

Why would it be any different to all the other code we've got behind
native pvops?

The ideal simplified case is that we rename
smp_send_stop/send_reschedule/prepare_cpus/etc to native_* versions.  In
the !PARAVIRT case we just call the native_* version directly; in
PARAVIRT we call via the native pv_ops structure.  Under Xen, all these
would implemented independently from the native versions.

> No, it's not the IPI itself, it's the way it's often accessed by the rest of
> the kernel (which is intertwined with genapic).  I'm happy to avoid apic
> altogether since it's effectively worthless for Xen other than
> integrating into the existing infrastructure.
>   

I guess by "rest of the kernel" you mean other stuff in arch/i386.  Yes,
that's a concern, but maybe we can tease it apart in a sensible way.

    J