[PATCH RFC] Introduce one machine_desc instance and get_machine_desc()

[Nicolas Pitre]

On Wed, 7 Jul 2010, Eric Miao wrote:

> commit 5d5d90e5d41bb0842559dd2b00f00f2a0f532a3a
> Author: Eric Miao <eric.miao at canonical.com>
> Date:   Wed Jul 7 16:47:20 2010 +0800
> 
>     [ARM] Introduce one machine_desc instance and get_machine_desc()
> 
>     If some of the fields remains useful in 'struct machine_desc', it's
>     normally a bit convenient to keep one instance there for reference.
> 
>     (along with more and more machine specific fields coming into this
>     structure, e.g. pcibios_min_{io,mem})
> 
>     It also improves modularity a bit by not exporting variables like
>     system_timer, init_arch_irq etc.
> 
>     However, the drawbacks are 1) a bit increased data and 2) increased
>     risk of accessing those fields which are marked as __init.

That makes me worried.  Access to __init data/functions after __init 
stuff is freed is amongst the nastiest subtle bugs that are hard to 
reproduce and therefore fix.  And copying such pointers is probably 
defeating the simple compile time checks we currently have against that.

Maybe those fields should be partitioned differently.  Having a new 
structure to encapsulate all fields that should be kept after __init 
stuff is gone might be a better way.


Nicolas