[Intel-gfx] [PATCH 4/4] drm/i915: Retry for live status

[Daniel Vetter]

On Thu, Aug 27, 2015 at 02:46:26PM +0530, Sharma, Shashank wrote:
> Regards
> Shashank
> 
> On 8/26/2015 8:47 PM, Daniel Vetter wrote:
> >On Wed, Aug 26, 2015 at 10:05:00AM +0000, Jindal, Sonika wrote:
> >>HPD bits control the interrupt but the live status (with some monitors) takes time to get set.
> >>We had experienced this with VLV and CHV with few monitors.
> >>So Android code always has this retry for live status.
> >>
> >>Yes, this was not added in the previous series because we planned to add the next set of optimization a little while later.
> >>But this seems to be an important one.
> >>
> >>It will be great if you can try it with your ivb. But for that you would need to first change the gen check and add a call to check live status for ivb.
> >
> >Done (well I just quickly hacked up the same idea on top of your old
> >patches). Lessons to be learned from this:
> >- Make sure that you really include _all_ the bugfixes. This pach here
> >   isn't just tuning, it's crucial to make it work. And this isn't the
> >   first time vpg teams upstream something and later on we notice that
> >   important bugfixes have been forgotten.
> >
> >   Because this wasn't done both you & me wasted a lot of time arguing
> >   about these patches and trying to test them.
> >
> Agree. We thought once the basic optimization goes in, we will add this as
> fine tuning patch. We were afraid of you guys doubting this approach at
> first itself. It looks like a little hack, but the HW itself is screwed up
> like this, to deal with.

Reality always wins, even if it looks really broken at first. The
important bit in those cases is to justify the change with a really good
commit message to make it clear what exactly is going on and why this
change is the correct fix/work-around.

> >- Please squash this patch in with patch 3 since otherwise we have a
> >   regression. Also please try to dig out why exactly this works like this
> >   since the hpd irq happening _before_ hpd status settles sounds to me
> >   like we have a little time machine in our silicon which can predict the
> >   future ...
> >
> Actually this depends on the monitor also. Few monitors are slow to assert
> the HPD line, or sometimes they don't provide the right voltage on that,
> causing live status to fluctuate for a while. While VLV/CHV beta testing we
> have done this experiment with a big range of monitors, and concluded that
> 30ms(retry of 10ms * 3) is the optimized time where most of the monitors
> respond well.
> 
> We saw that we cant delay further, because HDCP compliance expects us to
> respond to HPD (out) with in 100ms. So after careful testing with many
> monitors, we have concluded this range.

So what's happening on the wire is
                               
                              /------------
                ^      /\    /
               / \/\  /  \/\/              <- hpd threshold
----------/\--/     \/

                ^ first early hpd causing irq
                               ^ hpd status stable

If that's what's going on then yeah your patch makes a lot of sense, since
it perfectly explains why we've seen this on all platforms. Have you
confirmed this with oscilloscopes and all that? Please add all that
information (including pretty ascii graphs pls) to the commit message to
make sure we really have this documented correctly.
			  
> >- Please respin the patch series with the IS_VLV || gen >= 8 checks drop,
> >   I'm fairly confident that this bugfix here is the bit we've been looking
> >   for since years. At least it would be good to retest on all platforms
> >   for maximal test coverage.
> >
> Sure. Will do that.

Awesome, really looking forward to finally settling this after years!

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