On 2019-08-19 19:33:18 [-0500], Clark Williams wrote:
> From: Clark Williams <williams@xxxxxxxxxx>
>
> The following structures contain a member named 'irq_lock'.
> These three locks are of type spinlock_t and are used in
> multiple contexts including atomic:
>
> struct drm_i915_private
> struct intel_breadcrumbs
> strict intel_guc
>
> Convert them all to be raw_spinlock_t so that lockdep and the lock
> debugging code will be happy.
What is your motivation to make the lock raw?
I did the following:
void intel_engine_signal_breadcrumbs(struct intel_engine_cs *engine)
{
- local_irq_disable();
- intel_engine_breadcrumbs_irq(engine);
- local_irq_enable();
+ if (IS_ENABLED(CONFIG_PREEMPT_RT_FULL)) {
+ intel_engine_breadcrumbs_irq(engine);
+ } else {
+ local_irq_disable();
+ intel_engine_breadcrumbs_irq(engine);
+ local_irq_enable();
+ }
}
and lockdep was quiet (+ ignoring/patching the lockdep-irq-off-asserts).
The local_irq_disable() is here (my interpretation of the situation)
because that function is called from process context while the remaining
callers invoke intel_engine_breadcrumbs_irq() from the interrupt
handler and it acquires irq_lock via a plain spin_lock(). That
local_irq_disable() would be required if everyone did a _irqsave().
I tried to check how much worse the latency gets here but I didn't see
anything in a brief test. What I saw however is that switching to
fullscreen while playing a video gives me ~0.5 to ~2ms latency. This is
has nothing to do with this change, I have to dig deeper… It might be
one of the preempt_disable() section I just noticed.
I would prefer to keep the lock non-raw unless there is actual need for
it.
Sebastian
