-tip: ACPICA: Do not schedule during early init

[Ingo Molnar]

* Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote:

> Matt suggested, whereupon Linus sayeth:
> 
>  "This looks like a good patch. Please make it so"
> 
> Compile and boot tested on x86_64 only.

I wanted to wait with his for .32, but Linus applied it yesterday so 
lets hope it's all fine on all architectures ;-)

There's one new test failure on x86 caused by this commit: a 
scheduling-while-atomic assert during early ACPI init - fixed by the 
patch below. It triggers on two separate test-systems.

Btw., that ACPI_PREEMPTION_POINT() wrapper is both superfluous (we 
should not wrap something as obvious as a cond_resched()) and shows 
signs of problems:

#define ACPI_PREEMPTION_POINT() \
        do { \
                if (!irqs_disabled()) \
                        cond_resched(); \
        } while (0)

that should have been 1) an inline function 2) should not check for 
whether irqs are off. If we need to check for irqs-off like this 
then this is a sign that either the code flow is too unbalanced 
(mixing different things into the same function), or that the 
preemption point has been applied at a too low level.

So a followup cleanup would likely be in order, especially that this 
was the last (and only) call site of ACPI_PREEMPTION_POINT(). I'd 
suggest to remove it.

( I'm not sure what prompted the addition of this rescheduling point 
  - if there was a strong reason for it then we should probably add
  back the preemption point to the place that is causing the
  latency. )

	Ingo

-------------------->