Re: [RFC v5 12/12] accel: centralize initialization of CpusAccelOps

[Eduardo Habkost <ehabkost@redhat.com>]

On Tue, Nov 24, 2020 at 07:52:15PM +0100, Claudio Fontana wrote:
[...]
> >> +    }
> > 
> > Additionally, if you call arch_cpu_accel_init() here, you won't
> > need MODULE_INIT_ACCEL_CPU anymore.  The
> > 
> >   module_call_init(MODULE_INIT_ACCEL_CPU)
> > 
> > call with implicit dependencies on runtime state inside vl.c and
> > *-user/main.c becomes a trivial:
> > 
> >   accel_init(accel)
> > 
> > call in accel_init_machine() and *-user:main().
> 
> 
> 
> I do need a separate thing for the arch cpu accel specialization though,
> without MODULE_INIT_ACCEL_CPU that link between all operations done at accel-chosen time is missing..
> 

I think this is a key point we need to sort out.

What do you mean by "link between all operations done at
accel-chosen time" and why that's important?

accel_init_machine() has 2-3 lines of code with side effects.  It
calls AccelClass.init_machine(), which may may have hundreds of
lines of code.  accel_setup_post() has one additional method
call, which is triggered at a slightly different moment.

You are using MODULE_INIT_ACCEL for 2 additional lines of code:
- the cpus_register_accel() call
- the x86_cpu_accel_init() call

What makes those 2 lines of code so special, to make them deserve
a completely new mechanism to trigger them, instead of using
trivial function calls inside a accel_init() function?

-- 
Eduardo