Re: [PATCH v6 1/8] drm/msm/dp: Add eDP support via aux_bus

From: Abhinav Kumar
Date: Thu Apr 07 2022 - 18:09:34 EST


Hi Doug

Thanks for the response, some comments below.

Abhinav
On 4/7/2022 1:47 PM, Doug Anderson wrote:
Hi,

On Thu, Apr 7, 2022 at 1:11 PM Abhinav Kumar <quic_abhinavk@xxxxxxxxxxx> wrote:

Hi Doug and Dmitry

Sorry, but I caught up on this email just now.

Some comments below.

Thanks

Abhinav
On 4/7/2022 10:07 AM, Doug Anderson wrote:
Hi,

On Thu, Apr 7, 2022 at 7:19 AM Sankeerth Billakanti (QUIC)
<quic_sbillaka@xxxxxxxxxxx> wrote:

Hi Dmitry and Doug,

Hi,

On Tue, Apr 5, 2022 at 10:36 AM Dmitry Baryshkov
<dmitry.baryshkov@xxxxxxxxxx> wrote:

On 05/04/2022 20:02, Doug Anderson wrote:
Hi,

On Tue, Apr 5, 2022 at 5:54 AM Dmitry Baryshkov
<dmitry.baryshkov@xxxxxxxxxx> wrote:
3. For DP and eDP HPD means something a little different.
Essentially there are two concepts: a) is a display physically
connected and b) is the display powered up and ready. For DP, the
two are really tied together. From the kernel's point of view you
never "power down" a DP display and you can't detect that it's
physically connected until it's ready. Said another way, on you
tie "is a display there" to the HPD line and the moment a display
is there it's ready for you to do AUX transfers. For eDP, in the
lowest power state of a display it _won't_ assert its "HPD"
signal. However, it's still physically present. For eDP you simply
have to _assume_ it's present without any actual proof since you
can't get proof until you power it up. Thus for eDP, you report
that the display is there as soon as we're asked. We can't _talk_
to the display yet, though. So in get_modes() we need to be able
to power the display on enough to talk over the AUX channel to it.
As part of this, we wait for the signal named "HPD" which really means
"panel finished powering on" in this context.

NOTE: for aux transfer, we don't have the _display_ pipe and
clocks running. We only have enough stuff running to do the AUX
transfer.
We're not clocking out pixels. We haven't fully powered on the
display. The AUX transfer is designed to be something that can be
done early _before_ you turn on the display.


OK, so basically that was a longwinded way of saying: yes, we
could avoid the AUX transfer in probe, but we can't wait all the
way to enable. We have to be able to transfer in get_modes(). If
you think that's helpful I think it'd be a pretty easy patch to
write even if it would look a tad bit awkward IMO. Let me know if
you want me to post it up.

I think it would be a good idea. At least it will allow us to
judge, which is the more correct way.

I'm still happy to prototype this, but the more I think about it the
more it feels like a workaround for the Qualcomm driver. The eDP
panel driver is actually given a pointer to the AUX bus at probe
time. It's really weird to say that we can't do a transfer on it
yet... As you said, this is a little sideband bus. It should be able
to be used without all the full blown infra of the rest of the driver.

Yes, I have that feeling too. However I also have a feeling that just
powering up the PHY before the bus probe is ... a hack. There are no
obvious stopgaps for the driver not to power it down later.


Lets go back to why we need to power up the PHY before the bus probe.

We need to power up PHY before bus probe because panel-eDP tries to read
the EDID in probe() for the panel_id. Not get_modes().

So doug, I didnt follow your comment that panel-eDP only does EDID read
in get_modes()

panel_id = drm_edid_get_panel_id(panel->ddc);
if (!panel_id) {
dev_err(dev, "Couldn't identify panel via EDID\n");
ret = -EIO;
goto exit;
}

If we do not need this part, we really dont need to power up the PHY
before the probe(). The hack which dmitry was referring to.

Right. ...so we _could_ remove the above from the panel-edp probe and
defer it to get_modes() and it wouldn't be that hard. ...but:

1. It feels like a hack to work around the Qualcomm driver. The way
the AUX bus is designed is that a pointer to the AUX bus is passed to
the panel-edp probe. It seems kinda strange to say that the panel
isn't allowed to do transfers with the pointer that's passed in.


And thats why to satisfy the requirements of passing an initialized AUX, sankeerth is delaying devm_of_dp_aux_populate_ep_devices() till PHY is initialized which seems reasonable to satisfy the probe() time requirements.

Even if we move to pm_runtime(), yes I agree it will club all the resources needed to control AUX in one place but you will still have to initialize PHY before probe() under the hood of pm_runtime().

So how will it help this cause?

We just have to accept that initializing PHY is a requirement to use AUX and before calling panel-eDP's probe(), we have to have an initialized AUX.

So we are not working around the driver but just satisfying the hardware requirements to be able to satisfy panel-edp's and drm_panel_dp_aux_backlight()'s aux bus requirements.

2. There's a second place where we might do an AUX transfer at probe
time which is when we're using the DP AUX backlight. There we call
drm_panel_dp_aux_backlight(). Conceivably this too could be deferred
until the get_modes(), but now it feels even more like a hack. We're
going to be registering the backlight in the first call to
get_modes()? That's, ummm, unexpected. We could look at perhaps
breaking the "DP AUX backlight" in two parts also, but that gets
involved. I think we're supposed to know the number of backlight
levels at device init time for backlight devices and we need an AUX
transfer to that.




So the answer is that we could probably make it work, but it seems
like an uglier solution than just making the Qualcomm driver able to
do AUX transfers when it should be able to.

Correct and by delaying the panel-edp's probe(), we are doing exactly that?


So this is boiling down to why or how panel-eDP was originally designed.

This is why I think we need to move to Runtime PM to manage this. Basically:

1. When an AUX transfer happens, you grab a PM runtime reference that
_that_ powers up the PHY.

This will not be trivial and needs to be scoped out as sankeerth said
but if the above is the only concern, why do we need to do this? There
seems to be an explanation why we are doing this and its not a hack.

How would Dmitry's rework address this? We need some RFC to conclude on
that first.


2. At the end of the AUX transfer function, you do a "put_autosuspend".

Then it becomes not a hack, right?



pm runtime ops needs to be implemented for both eDP and DP. This change
take good amount of planning and code changes as it affects DP also.

Because this patch series consist of basic eDP changes for SC7280 bootup,
shall we take this pm_runtime implementation in subsequent patch series?

Dmitry is the real decision maker here, but in my opinion it would be
OK to get something landed first that worked OK and wasn't taking us
too far in the wrong direction and then we could get a follow up patch
to move to pm_runtime.

I would say the discussion changed into a direction of implementing
pm-runtime because the current patch series does what it takes to adhere
to panel-eDP's design along with aux bus requirements of PHY needing to
be on.

So doug, to answer your questions here:

"So I guess the net result is maybe we should just keep it where it is.
Long term I'd be interested in knowing if there's a reason why we
can't structure the driver so that AUX transfers can happen with less
intertwining with the rest of the code, but that can happen later. I
would expect that you'd basically just need clocks and regulators on
and maybe your PHY on."

Yes PHY needs to be absolutely on and configured before aux transfers.

If we want to change that up to stop reading the panel_id in the panel
probe() and do it later, perhaps some of the changes done here are not
needed.

It only seems reasonable that we first prototype that in a separate
patch even a RFC perhaps and take this further as these set of changes
are needed for basic display functionality on sc7280 chromebooks.

Let us know what are the concerns with doing it in a follow up change.

As per above, I'm not objecting to it being a follow-up change, but I
do believe it's the right design and will lead to an overall cleaner
solution. I think I even mentioned in my reviews that the current
patch series seems to "scattershot" enable resources and that's how we
end up with patches like patch #5 in this series ("drm/msm/dp: prevent
multiple votes for dp resources"). IMO there should be be a 1-to-1
mapping between "turn on resources" and "turn off resources" and it
should be reference counted. So if your codepath turned on resources
then it's up to your codepath to turn resources off when done. If a
seconde code path might be running at the same time then it should
also turn on/off resources itself. ...and it should all be managed by
pm_runtime which is _exactly designed_ for this specific use case.

Agreed on this topic, moving to pm_runtime() will club all resources in one place and make things cleaner that way.

Complexity of it obviously needs to be evaluated to check the effort Vs rewards.

But it will still not address the original concern of this thread that
powering up the PHY before the probe() is a hack.

"Yes, I have that feeling too. However I also have a feeling that just
> powering up the PHY before the bus probe is ... a hack. There are no
> obvious stopgaps for the driver not to power it down later."

We would still end up doing that under the hood of pm_runtime.

And thats why its an improvement and not a necessity qualifying it for a separate change.

-Doug