[PATCH 2/2] drm/amd/display: use drm_crtc_set_vblank_offdelay()

[Hamza Mahfooz]

On 7/9/24 06:09, Daniel Vetter wrote:
> On Tue, Jul 09, 2024 at 11:32:11AM +0200, Daniel Vetter wrote:
>> On Mon, Jul 08, 2024 at 04:29:07PM -0400, Hamza Mahfooz wrote:
>>> Hook up drm_crtc_set_vblank_offdelay() in amdgpu_dm, so that we can
>>> enable PSR more quickly for displays that support it.
>>>
>>> Signed-off-by: Hamza Mahfooz <hamza.mahfooz at amd.com>
>>> ---
>>>   .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 30 ++++++++++++++-----
>>>   1 file changed, 22 insertions(+), 8 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>>> index fdbc9b57a23d..ee6c31e9d3c4 100644
>>> --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>>> +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c
>>> @@ -8231,7 +8231,7 @@ static int amdgpu_dm_encoder_init(struct drm_device *dev,
>>>   
>>>   static void manage_dm_interrupts(struct amdgpu_device *adev,
>>>   				 struct amdgpu_crtc *acrtc,
>>> -				 bool enable)
>>> +				 struct dm_crtc_state *acrtc_state)
>>>   {
>>>   	/*
>>>   	 * We have no guarantee that the frontend index maps to the same
>>> @@ -8239,12 +8239,25 @@ static void manage_dm_interrupts(struct amdgpu_device *adev,
>>>   	 *
>>>   	 * TODO: Use a different interrupt or check DC itself for the mapping.
>>>   	 */
>>> -	int irq_type =
>>> -		amdgpu_display_crtc_idx_to_irq_type(
>>> -			adev,
>>> -			acrtc->crtc_id);
>>> +	int irq_type = amdgpu_display_crtc_idx_to_irq_type(adev,
>>> +							   acrtc->crtc_id);
>>> +	struct dc_crtc_timing *timing;
>>> +	int offdelay;
>>> +
>>> +	if (acrtc_state) {
>>> +		timing = &acrtc_state->stream->timing;
>>> +
>>> +		/* at least 2 frames */
>>> +		offdelay = 2000 / div64_u64(div64_u64((timing->pix_clk_100hz *
>>> +						       (uint64_t)100),
>>> +						      timing->v_total),
>>> +					    timing->h_total) + 1;
>>
>> Yeah, _especially_ when you have a this short timeout your really have to
>> instead fix the vblank driver code properly so you can enable
>> vblank_disable_immediate. This is just cheating :-)
> 
> Michel mentioned on irc that DC had immediate vblank disabling, but this
> was reverted with f64e6e0b6afe ("Revert "drm/amdgpu/display: set
> vblank_disable_immediate for DC"").
> 
> I haven't looked at the details of the bug report, but stuttering is
> exactly what happens when the driver's vblank code has these races. Going
> for a very low timeout instead of zero just means it's a bit harder to hit
> the issue, and much, much harder to debug properly.
> 
> So yeah even more reasons to look at the underlying root-cause here I
> think.
> -Sima

The issue is that DMUB (display firmware) isn't able to keep up with all of
the requests that the driver is making. The issue is fairly difficult to
reproduce (I've only seen it once after letting the system run with a
program that would engage PSR every so often, after several hours).
It is also worth noting that we have the same 2 idle frame wait on the 
windows
driver, for the same reasons. So, in all likelihood if it is your 
opinion that
the series should be NAKed, we will probably have to move the wait into the
driver as a workaround.

-- 
Hamza