Re: [PATCH] drm/i915/display: Optimize panel power-on wait time

[Jani Nikula <jani.nikula@linux.intel.com>]

On Tue, 01 Jul 2025, Lucas De Marchi <lucas.demarchi@intel.com> wrote:
> On Tue, Jul 01, 2025 at 12:28:41PM +0300, Jani Nikula wrote:
>>On Mon, 30 Jun 2025, Dibin Moolakadan Subrahmanian <dibin.moolakadan.subrahmanian@intel.com> wrote:
>>>  The current wait_panel_on() uses intel_de_wait() with a long timeout
>>>  (5000ms), which is suboptimal on Xe platforms where the underlying
>>>  xe_mmio_wait32() employs an exponential backoff strategy. This leads
>>>  to unnecessary delays during resume or power-on  when the panel becomes
>>>  ready earlier than the full timeout.
>>>
>>>  This patch splits the total wait time into two pases
>>>     - First wait for the typical panel-on time(180ms)
>>>     - If panel is not ready , continue polling in short 20ms intervals
>>>       until the maximum timeout (5000ms) is reached
>>
>>I'm *very* reluctant to merge any new custom wait hacks. I'm in the
>>process of *removing* a plethora of them [1][2][3].
>
> good riddance

Yay!

>>
>>I think the question is, should xe_mmio_wait32() (and the i915
>>counterpart) and the intel_de_wait*() functions be migrated to an
>>interface similar to read_poll_timeout(), i.e. provide sleep and timeout
>>separately, no exponential backoff. And implement the waits using
>>read_poll_timeout(), or whatever Ville ends up with [4].
>
> I saw your patch series and I'm eagerly waiting it to either propagate
> it in xe or have someone merge such a patch.  I'm not sure about
> removing the exponential backoff is a good thing overall, but if it's
> needed then it needs to be justified to add a new function to pair with
> read_poll_timeout(), not a custom driver function.

While I'm negative about the patch at hand, the underlying problem is
very real.

I think at the very least the exponential sleep backoff needs an upper
bound that's relative to the timeout. Maybe 10-25% of timeout?

With the example case of 5 second timeout, the exponential backoff
starting from 10 us leads to a dozen polls before reaching 100 ms
elapsed time, but then polls at approximately 1 s, 2 s, 4 s, and 8 s
elapsed time. Longer timeouts are of course rare, but they exhibit
increasingly worse behaviour.

So if what we're waiting takes 2.1 seconds, the next check will be about
2 seconds later. Similarly, if it takes 4.1 seconds, the next check will
be about 4 seconds later, in this case exceeding the timeout by 3
seconds.

Anyway, if xe_mmio_wait32() remains as it is, with read_poll_timeout()
it's trivial to do the wait in the intel_de_*() macros, in display side,
with sleeps and timeouts defined in display. Because for most things the
really quick fast polls are useless in display.


BR,
Jani.


-- 
Jani Nikula, Intel