[Intel-gfx] [PATCH 23/30] drm/i915: Cache GT fifo count for SandyBridge

Ben Widawsky ben at bwidawsk.net
Thu Apr 14 04:21:30 CEST 2011


On Tue, Apr 12, 2011 at 09:31:51PM +0100, Chris Wilson wrote:
> The read back of the available FIFO entries is vital for system
> stability, but extremely costly. However, we only need a guide so as to
> avoid eating into the reserved entries and since we are the only
> consumer we can cache the read of the count from the last write.
> 
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> ---
>  drivers/gpu/drm/i915/i915_drv.c |   14 +++++++++-----
>  drivers/gpu/drm/i915/i915_drv.h |    1 +
>  2 files changed, 10 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> index c416c1d..1146abd 100644
> --- a/drivers/gpu/drm/i915/i915_drv.c
> +++ b/drivers/gpu/drm/i915/i915_drv.c
> @@ -287,12 +287,16 @@ void __gen6_gt_force_wake_put(struct drm_i915_private *dev_priv)
>  
>  void __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
>  {
> -	int loop = 500;
> -	u32 fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
> -	while (fifo < 20 && loop--) {
> -		udelay(10);
> -		fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
> +	if (dev_priv->gt_fifo_count < 20 ) {
> +		int loop = 500;
> +		u32 fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
> +		while (fifo < 20 && loop--) {
> +			udelay(10);
> +			fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
> +		}
> +		dev_priv->gt_fifo_count = fifo;
>  	}
> +	dev_priv->gt_fifo_count--;
>  }
>  
>  static int i915_drm_freeze(struct drm_device *dev)
> diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
> index 2f45228..c837e10 100644
> --- a/drivers/gpu/drm/i915/i915_drv.h
> +++ b/drivers/gpu/drm/i915/i915_drv.h
> @@ -268,6 +268,7 @@ typedef struct drm_i915_private {
>  	int relative_constants_mode;
>  
>  	void __iomem *regs;
> +	u32 gt_fifo_count;
>  
>  	struct intel_gmbus {
>  		struct i2c_adapter adapter;

I'm sure you noticed that we have seriously problem both here and in the
put()/get() if the condition doesn't clear up in loop number of times.

I'd probably add a WARN(!loop, "uh oh"), but the patch is better than
what is there currently, so I'm okay either way.

Reviewed-by: Ben Widawsky <ben at bwidawsk.net>



More information about the Intel-gfx mailing list