Re: [patch/rfc 2.6.24-rc8-git] genirq: partial lockdep fixes

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2008-01-18 at 14:29 -0800, David Brownell wrote:
> EXPERIMENTAL and incomplete patch to make LOCKDEP behave better in
> the face of irq chaining, by annotating irqs with a lock_class which
> can reflect that hierarchy.

This is too short for me (who has no clue about genirq) to understand
what this patch does or why it does it.

> This version of the patch is incomplete in at least two respects:
> 
>  - There's no spin_lock_irq_nested() primitive, so that locking calls
>    on irq probing (yeech!) paths must ignore the annotations.
> 
> 	==> LOCKDEP feature is evidently missing:
> 		spin_lock_irq_nested(lock_ptr, lock_class)

This rant is more lines than adding the API :-/ the reason for it not
being there is simple, it wasn't needed up until now.

>  - We'd really need new API to declare the parent/child relationships
>    between IRQs to handle this properly.  Lacking such calls, this just
>    piggybacks set_irq_chained_handler() to modify the parent's class.
>    That's a problem, since only the lowest level of any lock tree gets
>    accurate nesting annotations.  On the plus side: no platform changes
>    are needed this way, so testing is easy.
> 
> 	==> GENIRQ programming interface evidently missing
> 		set_irq_parent(parent_irqnum, child_irqnum) (?)
> 
> Example:  when a GPIO controller's set_wake() handler needs to call its
> parent's set_irq_wake() method to ensure all appropriate signal paths are
> set up, that currently generates a bogus lockdep warning.  This patch
> removes those bogus warnings ... when there's only one level of parent.

OK.... still need a bit more detail here.