[Intel-gfx] [PATCH] drm/i915: Delegate our irq handler to a thread

[Tvrtko Ursulin]

On 26/09/2019 15:25, Chris Wilson wrote:
> Moving our primary irq handler to a RT thread incurs an extra 1us delay
> in process interrupts. This is most notice in waking up client threads,
> where it adds about 20% of extra latency. It also imposes a delay in
> feeding the GPU, an extra 1us before signaling secondary engines and
> extra latency in resubmitting work to keep the GPU busy. The latter case
> is insignificant as the latency hidden by the active GPU, and
> preempt-to-busy ensures that no extra latency is incurred for
> preemption.
> 
> The benefit is that we reduced the impact on the rest of the system, the
> cycletest shows a reduction from 5us mean latency to 2us, with the
> maximum observed latency (in a 2 minute window) reduced by over 160us.
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> Cc: Clark Williams <williams at redhat.com>
> Cc: Sebastian Andrzej Siewior <bigeasy at linutronix.de>
> ---
> Note this should need the fixes in
> 20190926105644.16703-2-bigeasy at linutronix.de, by itself this should be a
> test vehicle to exercise that patch!
> ---
>   drivers/gpu/drm/i915/i915_irq.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_irq.c b/drivers/gpu/drm/i915/i915_irq.c
> index bc83f094065a..f3df7714a3f3 100644
> --- a/drivers/gpu/drm/i915/i915_irq.c
> +++ b/drivers/gpu/drm/i915/i915_irq.c
> @@ -4491,8 +4491,8 @@ int intel_irq_install(struct drm_i915_private *dev_priv)
>   
>   	intel_irq_reset(dev_priv);
>   
> -	ret = request_irq(irq, intel_irq_handler(dev_priv),
> -			  IRQF_SHARED, DRIVER_NAME, dev_priv);
> +	ret = request_threaded_irq(irq, NULL, intel_irq_handler(dev_priv),
> +				   IRQF_SHARED, DRIVER_NAME, dev_priv);
>   	if (ret < 0) {
>   		dev_priv->drm.irq_enabled = false;
>   		return ret;
> 

Two questions:

1. Should we split out the master_ctl handling into the fast handler? 
Although can we pass anything from the fast to threaded handler? If not 
we'd have to re-read the master_ctl from the threaded handler.

2. What about our tasklets - with threaded irqs we don't need them any 
more, right? So in this case they just add additional latency.

Regards,

Tvrtko