[Intel-gfx] [PATCH] drm/i915: Reduce Data Link N value for 1 lane DP->hdmi converters

[Jani Nikula]

On Thu, 23 Mar 2017, Clint Taylor <clinton.a.taylor at intel.com> wrote:
> On 03/23/2017 05:30 AM, Jani Nikula wrote:
>> On Thu, 23 Mar 2017, clinton.a.taylor at intel.com wrote:
>>> From: Clint Taylor <clinton.a.taylor at intel.com>
>>>
>>> Several major vendor USB-C->HDMI converters fail to recover a 5.4 GHz 1 lane
>>> signal if the Data Link N is greater than 0x80000.
>>> Patch detects when 1 lane 5.4 GHz signal is being used and makes the maximum
>>> value 20 bit instead of the maximum specification supported 24 bit value.
>>>
>>> Cc: Jani Nikula <jani.nikula at intel.com>
>>> Cc: Anusha Srivatsa <anusha.srivatsa at intel.com>
>>>
>>
>> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=93578
>
> I will add to the commit message.
>
>>
>>> Signed-off-by: Clint Taylor <clinton.a.taylor at intel.com>
>>> ---
>>>  drivers/gpu/drm/i915/i915_reg.h      |    2 ++
>>>  drivers/gpu/drm/i915/intel_display.c |   15 +++++++++++----
>>>  2 files changed, 13 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
>>> index 04c8f69..838d8d5 100644
>>> --- a/drivers/gpu/drm/i915/i915_reg.h
>>> +++ b/drivers/gpu/drm/i915/i915_reg.h
>>> @@ -4869,6 +4869,8 @@ enum {
>>>
>>>  #define  DATA_LINK_M_N_MASK	(0xffffff)
>>>  #define  DATA_LINK_N_MAX	(0x800000)
>>> +/* Maximum N value useable on some DP->HDMI converters */
>>> +#define  DATA_LINK_REDUCED_N_MAX (0x80000)
>>>
>>>  #define _PIPEA_DATA_N_G4X	0x70054
>>>  #define _PIPEB_DATA_N_G4X	0x71054
>>> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
>>> index 010e5dd..6e1fdd2 100644
>>> --- a/drivers/gpu/drm/i915/intel_display.c
>>> +++ b/drivers/gpu/drm/i915/intel_display.c
>>> @@ -6315,9 +6315,10 @@ static int intel_crtc_compute_config(struct intel_crtc *crtc,
>>>  }
>>>
>>>  static void compute_m_n(unsigned int m, unsigned int n,
>>> -			uint32_t *ret_m, uint32_t *ret_n)
>>> +			uint32_t *ret_m, uint32_t *ret_n,
>>> +			uint32_t max_link_n)
>>>  {
>>> -	*ret_n = min_t(unsigned int, roundup_pow_of_two(n), DATA_LINK_N_MAX);
>>> +	*ret_n = min_t(unsigned int, roundup_pow_of_two(n), max_link_n);
>>
>> If there's evidence suggesting "certain other operating systems" always
>> use a max (or fixed value) of 0x80000, perhaps we should just follow
>> suit? Simpler and less magical.
>>
>
> The other OS's don't appear to be fixed to 0x80000. The calculation in 
> i915 rounds up to the nearest power of 2 and the other OS's might have a 
> slightly different calculation to the nearest power of 2. Of course I 
> haven't seen the other OS's code to know their exact formula. HBR3 will 
> cause a higher value to be calculated and having a fixed value may cause 
> issues. The i915 formula works and reducing the value can cause 
> precision issues in the ratio with the pixel clock.
>
>>>  	*ret_m = div_u64((uint64_t) m * *ret_n, n);
>>>  	intel_reduce_m_n_ratio(ret_m, ret_n);
>>>  }
>>> @@ -6327,14 +6328,20 @@ static void compute_m_n(unsigned int m, unsigned int n,
>>>  		       int pixel_clock, int link_clock,
>>>  		       struct intel_link_m_n *m_n)
>>>  {
>>> +	uint32_t max_link_n = DATA_LINK_N_MAX;
>>>  	m_n->tu = 64;
>>>
>>> +	if ((nlanes==1) && (link_clock >= 540000))
>>
>> Is the problem really dependent on these conditions? You can get the
>> same problematic N value with nlanes == 2 && link_clock == 270000 too.
>>
>
> The offending device only supports a single DP lane up to HBR2.5. This 
> check matches the datasheet for the part. The offending device works 
> with our current calculation at 1 lane HBR (270000).

Okay, so what bugs me about the approach here is that this adds an
arbitrary condition to apply a quirk to a specific device.

Instead of "if device X, then apply restriction A", this adds "if
condition Y, then apply restriction A". If I understand you correctly,
"condition Y" is a superset of "device X", i.e. Y happens also on
devices other than X, but on device X condition Y always holds.

I'd really like it if we could come up with a) a quirk that we apply
only on the affected device(s), or b) rules for M/N that generally make
sense with no need to resort to seeminly arbitrary exceptions.

With the latter I mean things like reducing the M/N before rounding N up
to power of two (M and N are always divisible by 2, for example) or
having intel_reduce_m_n_ratio() shift them right as long as they have
bit 0 unset. At a glance, I'm not sure if this is enough to bring down
the N to within the limits of the device, without intentional loss of
precision.

BR,
Jani.


>
>> BR,
>> Jani.
>>
>>> +		max_link_n = DATA_LINK_REDUCED_N_MAX;
>>> +
>>>  	compute_m_n(bits_per_pixel * pixel_clock,
>>>  		    link_clock * nlanes * 8,
>>> -		    &m_n->gmch_m, &m_n->gmch_n);
>>> +		    &m_n->gmch_m, &m_n->gmch_n,
>>> +		    max_link_n);
>>>
>>>  	compute_m_n(pixel_clock, link_clock,
>>> -		    &m_n->link_m, &m_n->link_n);
>>> +		    &m_n->link_m, &m_n->link_n,
>>> +		    max_link_n);
>>>  }
>>>
>>>  static inline bool intel_panel_use_ssc(struct drm_i915_private *dev_priv)
>>
>

-- 
Jani Nikula, Intel Open Source Technology Center