[PATCH 2/2] ARM: SMP: use per cpu state to replace

[Thomas Gleixner]

On Mon, 21 May 2012, Russell King - ARM Linux wrote:

> On Mon, May 21, 2012 at 05:45:31PM +0800, Hui Wang wrote:
> > CPU hotplug will bring following call trace in the RT kernel:
> > BUG: sleeping function called from invalid context at linux/kernel/rtmutex.c:707
> > pcnt: 1 0 in_atomic(): 1, irqs_disabled(): 128, pid: 0, name: swapper
> > [<800413a0>] (unwind_backtrace+0x0/0xe4) from [<804ff214>] (rt_spin_lock+0x30/0x5c)
> > [<804ff214>] (rt_spin_lock+0x30/0x5c) from [<8005a848>] (complete+0x1c/0x54)
> > [<8005a848>] (complete+0x1c/0x54) from [<804f59f8>] (cpu_die+0x34/0x70)
> > [<804f59f8>] (cpu_die+0x34/0x70) from [<8003b840>] (cpu_idle+0x54/0xd8)
> > [<8003b840>] (cpu_idle+0x54/0xd8) from [<104f9ecc>] (0x104f9ecc)
> > 
> > To avoid this call trace, we use per cpu variable to replace
> > completion, and it is safe for this modification since all reference
> > of per cpu_state variable is in the preempt disabled context.
> 
> This is silly.  Why is RT preventing things that work perfectly well in
> the standard kernel from being used in the RT kernel?

The silliness is in some of the users of waitqueues. waitqueues are
protected by a spinlock, which we cannot convert to a raw spinlock on
rt because some of the wq users run insane callbacks from the wakeup
context with interrupts disabled.....
 
> Being able to call complete() from atomic contexts is one of the
> fundamentals that RT seems to be breaking here.

Yeah, I have a simpler waitqueue variant (no callbacks) which I want
to use for completions.

Thanks,

	tglx