[Intel-gfx] BUG: sleeping function called from invalid context on 3.10.10-rt7

[Steven Rostedt]

On Wed, 25 Sep 2013 06:32:10 +0200
Mario Kleiner <mario.kleiner at tuebingen.mpg.de> wrote:


> I assume if a spin_lock_irqsave doesn't really disable interrupts on a 
> RT kernel with normal spinlocks then local_irq_disable won't really 
> disable interrupts either?
> 

That is incorrect. On PREEMPT_RT, you are right about
spin_lock_irqsave() not disabling interrupts, but local_irq_disable()
does indeed disable interrupts.

Open coded local_irq_disable() usually ends up being a bug as it does
nothing to synchronize interrupts from other CPUs. But most of those
bugs have been removed, and there are some very legit reasons for using
local_irq_disable(). PREEMPT_RT honors those.

The reason PREEMPT_RT does not disable interrupts for
spin_lock_irqsave(), is because that's usually used for when a
interrupt uses the same spinlock. You need the irqsave() part in order
to prevent a deadlock, if the code that has the spinlock gets preempted
by the interrupt and that interrupt tries to grab the same lock.

Because PREEMPT_RT runs interrupts as threads, we don't have that issue,
because if the interrupt preempts the holder of the lock and it tries
to grab the same lock, it will just block like any other thread trying
to grab that lock. That is, spinlocks turn into mutexes on PREEMPT_RT.

-- Steve