Re: acpi_evaluate_integer broken by design

[Len Brown <lenb@kernel.org>]

> > Now I know why I had strange "scheduling in atomic" problems:
> > acpi_evaluate_integer() does malloc(..., irqs_disabled() ? GFP_ATOMIC
> > : GFP_KERNEL)... which is (of course) broken.
> 
> That is kinda weird.  When did this all start happening?

> > There's no way to reliably tell if we need GFP_ATOMIC or not from
> > code, this one for example fails to detect spinlocks held.

> Len, this looks like 2.6.28 material.  But given the poor quality of
> the changelog it is hard to be sure about this.  Why isn't everyone
> seeing these warnings?  What did Pavel do to provoke these alleged
> warnings?  Nobody knows...

I don't know know why pavel sees this and nobody else --
maybe something unusual he's doing with suspend?

The reason that the ACPI code is littered with bogus
irqs_disabled() ? GFP_ATOMIC : GFP_KERNEL)
is because, like boot, resume starts life with interrupts off.

I would prefer that resume and boot handle this the same way,
with system_state.  However, a few years ago when I suggested
using system_state for resume, Andrew thought that was a very
bad idea.  Andrew, do you still feel that way?

-Len

ps. I'll put this particular fix in my tree now.