Chris Wilson <[email protected]> writes:
> In the idle worker we drop the prolonged GT wakeref used to cover such
> essentials as interrupt delivery. (When a CS interrupt arrives, we also
> assert that the GT is awake.) However, it turns out that 10ms is not
> long enough to be assured that the last CS interrupt has been delivered,
> so bump that to 200ms, and move the entirety of that wait to before we
> take the struct_mutex to avoid blocking. As this is now a potentially
> long wait, restore the earlier behaviour of bailing out early when a new
> request arrives.
>
> Signed-off-by: Chris Wilson <[email protected]>
> Cc: Joonas Lahtinen <[email protected]>
> Cc: Mika Kuoppala <[email protected]>
> Cc: Imre Deak <[email protected]>
> ---
> drivers/gpu/drm/i915/i915_gem.c | 31 ++++++++++++++++++++-----------
> 1 file changed, 20 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 026cb52ece0b..d3a638613857 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -3281,8 +3281,8 @@ i915_gem_idle_work_handler(struct work_struct *work)
> {
> struct drm_i915_private *dev_priv =
> container_of(work, typeof(*dev_priv), gt.idle_work.work);
> - struct drm_device *dev = &dev_priv->drm;
> bool rearm_hangcheck;
> + ktime_t end;
>
> if (!READ_ONCE(dev_priv->gt.awake))
> return;
> @@ -3291,14 +3291,22 @@ i915_gem_idle_work_handler(struct work_struct *work)
> * Wait for last execlists context complete, but bail out in case a
> * new request is submitted.
> */
> - wait_for(intel_engines_are_idle(dev_priv), 10);
> - if (READ_ONCE(dev_priv->gt.active_requests))
> - return;
> + end = ktime_add_ms(ktime_get(), 200);
> + do {
> + if (READ_ONCE(dev_priv->gt.active_requests) ||
> + work_pending(work))
> + return;
> +
> + if (intel_engines_are_idle(dev_priv))
> + break;
> +
> + usleep_range(100, 500);
> + } while (ktime_before(ktime_get(), end));
>
Why can't we just wait_for(intel_engines_are_idle(dev_priv, 200)) ?
-Mika
> rearm_hangcheck =
> cancel_delayed_work_sync(&dev_priv->gpu_error.hangcheck_work);
>
> - if (!mutex_trylock(&dev->struct_mutex)) {
> + if (!mutex_trylock(&dev_priv->drm.struct_mutex)) {
> /* Currently busy, come back later */
> mod_delayed_work(dev_priv->wq,
> &dev_priv->gt.idle_work,
> @@ -3310,13 +3318,14 @@ i915_gem_idle_work_handler(struct work_struct *work)
> * New request retired after this work handler started, extend active
> * period until next instance of the work.
> */
> - if (work_pending(work))
> + if (dev_priv->gt.active_requests || work_pending(work))
> goto out_unlock;
>
> - if (dev_priv->gt.active_requests)
> - goto out_unlock;
> -
> - if (wait_for(intel_engines_are_idle(dev_priv), 10))
> + /*
> + * We are committed now to parking the engines, make sure there
> + * will be no more interrupts arriving later.
> + */
> + if (!intel_engines_are_idle(dev_priv))
> DRM_ERROR("Timeout waiting for engines to idle\n");
>
> intel_engines_mark_idle(dev_priv);
> @@ -3330,7 +3339,7 @@ i915_gem_idle_work_handler(struct work_struct *work)
> gen6_rps_idle(dev_priv);
> intel_runtime_pm_put(dev_priv);
> out_unlock:
> - mutex_unlock(&dev->struct_mutex);
> + mutex_unlock(&dev_priv->drm.struct_mutex);
>
> out_rearm:
> if (rearm_hangcheck) {
> --
> 2.15.0.rc1
_______________________________________________
Intel-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
