[Intel-gfx] [PATCH v4 6/6] drm/i915: Add skl_check_nv12_surface for NV12

[Maarten Lankhorst]

Op 19-04-18 om 13:32 schreef Ville Syrjälä:
> On Thu, Apr 19, 2018 at 10:12:56AM +0200, Maarten Lankhorst wrote:
>> Op 18-04-18 om 20:35 schreef Ville Syrjälä:
>>> On Wed, Apr 18, 2018 at 08:06:57PM +0200, Maarten Lankhorst wrote:
>>>> Op 18-04-18 om 17:32 schreef Ville Syrjälä:
>>>>> On Wed, Apr 18, 2018 at 09:38:13AM +0530, Vidya Srinivas wrote:
>>>>>> From: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
>>>>>>
>>>>>> We skip src trunction/adjustments for
>>>>>> NV12 case and handle the sizes directly.
>>>>>> Without this, pipe fifo underruns are seen on APL/KBL.
>>>>>>
>>>>>> v2: For NV12, making the src coordinates multiplier of 4
>>>>>>
>>>>>> v3: Moving all the src coords handling code for NV12
>>>>>> to skl_check_nv12_surface
>>>>>>
>>>>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
>>>>>> Signed-off-by: Vidya Srinivas <vidya.srinivas at intel.com>
>>>>>> ---
>>>>>>  drivers/gpu/drm/i915/intel_display.c | 39 ++++++++++++++++++++++++++++++++++++
>>>>>>  drivers/gpu/drm/i915/intel_sprite.c  | 15 ++++++++++----
>>>>>>  2 files changed, 50 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
>>>>>> index 925402e..b8dbaca 100644
>>>>>> --- a/drivers/gpu/drm/i915/intel_display.c
>>>>>> +++ b/drivers/gpu/drm/i915/intel_display.c
>>>>>> @@ -3118,6 +3118,42 @@ static int skl_check_main_surface(const struct intel_crtc_state *crtc_state,
>>>>>>  	return 0;
>>>>>>  }
>>>>>>  
>>>>>> +static int
>>>>>> +skl_check_nv12_surface(const struct intel_crtc_state *crtc_state,
>>>>>> +		       struct intel_plane_state *plane_state)
>>>>>> +{
>>>>>> +	int crtc_x2 = plane_state->base.crtc_x + plane_state->base.crtc_w;
>>>>>> +	int crtc_y2 = plane_state->base.crtc_y + plane_state->base.crtc_h;
>>>>>> +
>>>>>> +	if (((plane_state->base.src_x >> 16) % 4) != 0 ||
>>>>>> +	    ((plane_state->base.src_y >> 16) % 4) != 0 ||
>>>>>> +	    ((plane_state->base.src_w >> 16) % 4) != 0 ||
>>>>>> +	    ((plane_state->base.src_h >> 16) % 4) != 0) {
>>>>>> +		DRM_DEBUG_KMS("src coords must be multiple of 4 for NV12\n");
>>>>>> +		return -EINVAL;
>>>>>> +	}
>>>>> I don't really see why we should check these. The clipped coordinates
>>>>> are what matters.
>>>> To propagate our limits to the userspace. I think we should do it for all formats,
>>>> but NV12 is the first YUV format we have tests for. If we could we should do
>>>> something similar for the other YUV formats, but they have different requirements.
>>>>
>>>> In case of NV12 we don't have existing userspace, there will be nothing that
>>>> breaks if we enforce limits from the start.
>>> But what about sub-pixel coordinates? You're totally ignoring them here.
>>> We need to come up with some proper rules for this stuff.
>> Would we break anything if we disallow sub-pixel coordinates for i915 globally? It's not like we supported them before,
>> but I'm not sure that change would break anything.
> Not really I suppose. IIRC the hw did reintroduce partial sub-pixel
> coordinate support for NV12 specifically. I do wish they'd done it
> fully for all formats.
>
>>>>>> +
>>>>>> +	/* Clipping would cause a 1-3 pixel gap at the edge of the screen? */
>>>>>> +	if ((crtc_x2 > crtc_state->pipe_src_w && crtc_state->pipe_src_w % 4) ||
>>>>>> +	    (crtc_y2 > crtc_state->pipe_src_h && crtc_state->pipe_src_h % 4)) {
>>>>>> +		DRM_DEBUG_KMS("It's not possible to clip %u,%u to %u,%u\n",
>>>>>> +			      crtc_x2, crtc_y2,
>>>>>> +			      crtc_state->pipe_src_w, crtc_state->pipe_src_h);
>>>>>> +		return -EINVAL;
>>>>>> +	}
>>>>> Why should we care? The current code already plays it fast and loose
>>>>> and allows the dst rectangle to shrink to accomodate the hw limits.
>>>>> If we want to change that we should change it universally.
>>>> Unfortunately for the other formats we already have an existing userspace
>>>> (X.org) that doesn't perform any validation. We can't change it for that,
>>>> but we can prevent future mistakes.
>>> We should do it uniformly. Not per-format. That will make the code
>>> unmaintainable real quick.
>>>>>> +
>>>>>> +	plane_state->base.src.x1 =
>>>>>> +		DIV_ROUND_CLOSEST(plane_state->base.src.x1, 1 << 18) << 18;
>>>>>> +	plane_state->base.src.x2 =
>>>>>> +		DIV_ROUND_CLOSEST(plane_state->base.src.x2, 1 << 18) << 18;
>>>>>> +	plane_state->base.src.y1 =
>>>>>> +		DIV_ROUND_CLOSEST(plane_state->base.src.y1, 1 << 18) << 18;
>>>>>> +	plane_state->base.src.y2 =
>>>>>> +		DIV_ROUND_CLOSEST(plane_state->base.src.y2, 1 << 18) << 18;
>>>>> Since this can now increase the size of the source rectangle our
>>>>> scaling factor checks are no longer 100% valid. We might end up with
>>>>> a scaling factor that is too high.
>>>>>
>>>>> I don't really like any of these "let's make NV12 behave special"
>>>>> tricks. We should make the code behave the same way for all pixel
>>>>> formats instead of adding format specific hacks.
>>>> This is not nivalid because we restrict the original src coordinates to be
>>>> a multiple of 4, you can only clip to something smaller, not to something
>>>> bigger. :)
>>> The clipped coordinates can be whatever thanks to scaling/etc.
>> Yes, but it will always be smaller than the original rectangle, so rounding to 4 when
>> the original set of coordinates were a multiple of 4 would never go outside the original
>> boundary.
> I was talking about the scaling factor increasing, and potentially
> exceeding the hardware maximum.
>
>>> Also why are we trying to make everything a multiple of four? I don't
>>> remember any hw restrictions like that.
>> Well Vidya already replied, it sucks but it's what we have to live with for now. :(
> That was just about the plane height. Nothing seems to require making
> everything a multiple of four.
>
This was to get rid of FIFO underruns, but the new solution appears to be not enable it on BXT. I can live with that. :)