[Intel-gfx] [PATCH] drm/i915: Skip waking the device to service pwrite

[Daniel Vetter]

On Wed, Aug 30, 2017 at 06:48:19PM +0100, Chris Wilson wrote:
> If the device is in runtime suspend, resuming takes time and reduces our
> powersaving. If this was for a small write into an object, that resume
> will take longer than any savings in using the indirect GGTT access to
> avoid the cpu cache.
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/i915_gem.c | 21 ++++++++++++++++++---
>  1 file changed, 18 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
> index 93dfa793975a..8940a6873ca5 100644
> --- a/drivers/gpu/drm/i915/i915_gem.c
> +++ b/drivers/gpu/drm/i915/i915_gem.c
> @@ -1229,7 +1229,21 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
>  	if (ret)
>  		return ret;
>  
> -	intel_runtime_pm_get(i915);
> +	if (i915_gem_object_has_struct_page(obj)) {

I don't really see why we need to check for has_struct_page here (we do
already outside the lock grabbing), and why if that's not the case we hit
the slow-path?

I'd have expected a simple s/pm_get/pm_get_if_in_use/ ...
-Daniel

> +		/* Avoid waking the device up if we can fallback, as
> +		 * waking/resuming is very slow (10-100 ms depending
> +		 * on PCI sleeps and our own resume time). This easily
> +		 * dwarfs any performance advantage from using the
> +		 * cache bypass of indirect GGTT access.
> +		 */
> +		if (!intel_runtime_pm_get_if_in_use(i915)) {
> +			ret = -EFAULT;
> +			goto out_unlock;
> +		}
> +	} else {
> +		intel_runtime_pm_get(i915);
> +	}
> +
>  	vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0,
>  				       PIN_MAPPABLE | PIN_NONBLOCK);
>  	if (!IS_ERR(vma)) {
> @@ -1244,7 +1258,7 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
>  	if (IS_ERR(vma)) {
>  		ret = insert_mappable_node(ggtt, &node, PAGE_SIZE);
>  		if (ret)
> -			goto out_unlock;
> +			goto out_rpm;
>  		GEM_BUG_ON(!node.allocated);
>  	}
>  
> @@ -1307,8 +1321,9 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj,
>  	} else {
>  		i915_vma_unpin(vma);
>  	}
> -out_unlock:
> +out_rpm:
>  	intel_runtime_pm_put(i915);
> +out_unlock:
>  	mutex_unlock(&i915->drm.struct_mutex);
>  	return ret;
>  }
> -- 
> 2.14.1
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch