Re: [RESEND PATCH] of: property: do not create clocks device link for clock controllers

[Dmitry Baryshkov <dmitry.baryshkov@linaro.org>]

On Fri, 27 Jan 2023 at 01:12, Saravana Kannan <saravanak@google.com> wrote:
>
> On Thu, Jan 26, 2023 at 2:51 PM Dmitry Baryshkov
> <dmitry.baryshkov@linaro.org> wrote:
> >
> > On 25/01/2023 21:09, Rob Herring wrote:
> > > On Tue, Jan 24, 2023 at 06:12:15PM -0800, Saravana Kannan wrote:
> > >> On Wed, Jan 18, 2023 at 5:35 AM Rob Herring <robh+dt@kernel.org> wrote:
> > >>>
> > >>> On Wed, Jan 18, 2023 at 3:11 AM Dmitry Baryshkov
> > >>> <dmitry.baryshkov@linaro.org> wrote:
> > >>>>
> > >>>> Do not create device link for clock controllers. Some of the clocks
> > >>>> provided to the device via OF can be the clocks that are just parents to
> > >>>> the clocks provided by this clock controller. Clock subsystem already
> > >>>> has support for handling missing clock parents correctly (clock
> > >>>> orphans). Later when the parent clock is registered, clocks get
> > >>>> populated properly.
> > >>>>
> > >>>> An example of the system where this matters is the SDM8450 MTP board
> > >>>> (see arch/arm64/boot/dts/qcom/sdm845-mtp.dts). Here the dispcc uses
> > >>>> clocks provided by dsi0_phy and dsi1_phy device tree nodes. However the
> > >>>> dispcc itself provides clocks to both PHYs, to the PHY parent device,
> > >>>> etc. With just dsi0_phy in place devlink is able to break the
> > >>>> dependency, but with two PHYs, dispcc doesn't get probed at all, thus
> > >>>> breaking display support.
> > >>>>
> > >>>> Cc: Bjorn Andersson <andersson@kernel.org>
> > >>>> Cc: Stephen Boyd <sboyd@kernel.org>
> > >>>> Cc: Saravana Kannan <saravanak@google.com>
> > >>>> Cc: Abel Vesa <abel.vesa@linaro.org>
> > >>>> Signed-off-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org>
> > >>>> ---
> > >>>>
> > >>>> This patch has been posted a year ago in January 2022 ([1]). Since that time
> > >>>> Saravana failed to produce patches to assist in debugging the issue
> > >>>> ([2]) or to fix the issue ([3]). The issue we observe has been described
> > >>>> by Abel at ([4]). As we work on adding support for Dual DSI
> > >>>> configurations, the issue becomes more and more important, since binding
> > >>>> the whole display subsystem fails.
> > >>
> > >> I did send out a patch series[1] to try and fix this. Heck I even
> > >> talked about this in LPC 2022. So I don't think it's accurate to say I
> > >> didn't help debug this or fix this. There's some email thread in lore
> > >> where Abel gave more details and I figured out the issue and we didn't
> > >> need any more debugging. And then I sent out [1]. Sorry I missed you
> > >> in the cc lise for [1] -- I try to keep track of everyone to cc but
> > >> things slip through the cracks sometimes. But at the same time, it's
> > >> easy to check for emails from me before saying I didn't help or didn't
> > >> send out fixes :)
> > >>
> > >> If you do try to give [1] a shot, there are a bunch of bugs that
> > >> people pointed out for which I gave fixes on top of [1] in the
> > >> replies. I was supposed to work on v2 over the holidays, but that
> > >> didn't happen because of stuff outside my control.
> > >>
> > >>> That's ample time to fix this, so I intend to apply this. But I'll
> > >>> give it a few days for comments.
> > >>
> > >> Rob, I'd recommend not applying this because it'll fix it for Dmitry
> > >> but break someone else's use case. That's the whole reason it takes me
> > >> a while to send out patches -- it's easy to fix it for a subset of
> > >> devices, but fixing something without breaking someone else is harder
> > >> (I still believe it's doable) and it takes a while to test them on all
> > >> the devices I want to test before sending them out.
> >
> > This case is really simple, I think. Clock controllers (and
> > clock-core-framework) are prepared to handle clock orphans properly.
> > Moreover they have been supposed to work in such way for quite a while.
> > In other words, I don't think we should save them from this
> > -EPROBE_DEFERRED.
>
> A clock controller can depend on other clock controllers for non clock
> tree reasons. For example, it might need a clock ON to access its
> registers. So, while the CCF can handle orphans properly, that's not
> the only dependency. Also, fw_devlink is not just about probing
> either. It also has to do with proper sync_state() callbacks.

Just a question, please excuse if I'm misunderstanding it. Does
fw_devlink created this way also impose any runtime PM dependencies?

>
> Also, I already fixed the issue you are referring to while not
> breaking the conditions I'm referring to. So, I don't know why you are
> so opposed to that. See Abel's Tested-by here:
> https://lore.kernel.org/lkml/YvonlAwXAoXTUTZe@linaro.org/
>
> > Thus I think it is better to let them continue doing their job of
> > handling probe deferrals on their own, at least for the time being.
>
> I'm pretty sure your patch will break other Qualcomm platforms because
> they depend on sync_state() callbacks to boot up properly when
> all/most of their drivers are built as modules.

Qualcomm platforms did not use sync state for clock controllers. Only
for the icc drivers.

>
> > And
> > then, when your patches are finished, we can think about reenabling
> > current behaviour. As a reminder, currently, all Qualcomm platforms
> > trying to use double DSI configuration are broken and have to use
> > fw_devlink= kernel params.
>
> I'm/was working on sending out the v2 when I got your email. Hold
> tight please. It shouldn't take too long.

I'll give v2 a test next week, thank you!

-- 
With best wishes
Dmitry