Re: [PATCH v2 0/3] QCM2290 LMH

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 26/03/2024 00:01, Nícolas F. R. A. Prado wrote:
> On Mon, Mar 25, 2024 at 08:59:55PM +0100, Krzysztof Kozlowski wrote:
>> On 20/03/2024 20:08, Nícolas F. R. A. Prado wrote:
>>>> Loic Poulain (1):
>>>>       arm64: dts: qcom: qcm2290: Add LMH node
>>>>
>>>>  Documentation/devicetree/bindings/thermal/qcom-lmh.yaml | 12 ++++++++----
>>>>  arch/arm64/boot/dts/qcom/qcm2290.dtsi                   | 14 +++++++++++++-
>>>>  drivers/thermal/qcom/lmh.c                              |  3 +++
>>>>  3 files changed, 24 insertions(+), 5 deletions(-)
>>>
>>> Hi,
>>>
>>> I've started tracking the results of 'make dtbs_check' on linux-next, and I've
>>> noticed that on today's next, next-20240320, there's a new warning coming from
>>> this. The reason is that the DT change has landed, but the binding has not,
>>> since it goes through a separate tree. I thought the binding was supposed to
>>> always land before the driver and DT that make use of it, but looking through
>>
>> There is no such rule. Of course new binding should be documented in
>> earlier or the same kernel release cycle as users get in, but it's not a
>> requirement.
> 
> So, after giving the documentation a second look, I found this:
> 
> "For new platforms, or additions to existing ones, make dtbs_check should not
> add any new warnings."
> 
> Source: https://www.kernel.org/doc/html/latest/process/maintainer-soc.html#validating-devicetree-files

It's just "should"...

> 
> What is not clear there is what the reference point is: is it on linux-next?
> Mainline release?

Does it matter? There was never a new warning introduced by this
patchset. The patchset itself is correct. No new warnings.

> 
> As Konrad pointed out it's tricky (and maybe not worth it) to guarantee this for
> linux-next. But for mainline release it seems feasible (and IMO the target, as
> after that stability guarantees should apply).

I don't believe in such guarantees. Different maintainers apply patches
differently, especially bindings, so this is beyond our control. Often
also beyond SoC maintainer control.

> 
>>
>>
>>> the dt-binding documentation pages I couldn't find anything confirming or
>>> denying that.
>>>
>>> I expect this to happen again in the future, which is why I'm reaching out to
>>> understand better how to deal with this kind of situation.
>>
>> Deal as what to do? Are you asking in terms of maintenance of some
>> subsystem or sending some patches? In this particular case here, I don't
>> think there is anything on your side to deal with.
> 
> I'm asking what's the most helpful way to you the maintainers for me to report
> these failures in the future.

The most effective way is LKP-like or Rob's-bot-like automated replies
to original email threads, by testing the original patchset on
linux-next. But Rob's bot is actually doing it, just on different base.

Other reports, like for cases when only parts of patch is applied, could
be also useful but I am afraid you will generate way too much of them.
Binding is supposed to go via subsystem, DTS via SoC, so basically 90%
of patchsets might have some sort of delays resulting in dtbs_check
false positive warnings.

For my SoC I check my trees, mainline and next, and keep adding list of
exceptions for expected issues. What's useful for Qualcomm? Konrad,
Bjorn, any thoughts?

Have in mind that expected warnings can be for entire cycle when dealing
with technical debt, because DTS goes N+1.

> 
> Rob has already automated running dtbs_check for patches coming into the mailing
> list. And I have set up KernelCI to run dtbs_check on linux-next in order to
> catch any issues that might slip through, or happen during integration of the
> trees, etc.
> 
> Now, if we agree that dtbs_check regressions on linux-next are acceptable, at
> least ones like this, where the issue is just synchronization between

Yes and no. True regressions are not acceptable. Expected intermediate
regressions as a result of patchset being applying, but not yet fully
applied, are OK. Expected regressions for intra-cycle-work are also OK.

> maintainers, then I can simply not report them in the future. But we should
> have some point where dtbs_check should not regress, and mainline release seems
> the reasonable choice, because if we don't then dtbs_check warnings would just
> keep growing forever.

I invite therefore to my session:
https://eoss24.sched.com/event/1aBEf?iframe=no
We'll see if they keep growing :)

Best regards,
Krzysztof





[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [Linux for Sparc]     [IETF Annouce]     [Security]     [Bugtraq]     [Linux MIPS]     [ECOS]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux