Re: [PATCH v4 00/16] R-Car DU: Convert LVDS code to bridge driver

[Frank Rowand]

On 02/23/18 01:25, Laurent Pinchart wrote:
> Hi Frank,
> 
> On Friday, 23 February 2018 05:20:43 EET Frank Rowand wrote:
>> On 02/22/18 02:25, Laurent Pinchart wrote:
>>> On Thursday, 22 February 2018 08:07:14 EET Frank Rowand wrote:
>>>> On 02/20/18 15:10, Laurent Pinchart wrote:
>>>>> Hello,
>>>>>
>>>>> This patch series addresses a design mistake that dates back from the
>>>>> initial DU support. Support for the LVDS encoders, which are IP cores
>>>>> separate from the DU, was bundled in the DU driver. Worse, both the DU
>>>>> and LVDS were described through a single DT node.
>>>>>
>>>>> To fix the, patches 01/16 and 02/16 define new DT bindings for the LVDS
>>>>> encoders, and deprecate their description inside the DU bindings. To
>>>>> retain backward compatibility with existing DT, patches 03/16 to 08/16
>>>>> then patch the device tree at runtime to convert the legacy bindings to
>>>>> the new ones.
>>>>>
>>>>> With the DT side addressed, patch 09/16 converts the LVDS support code
>>>>> to a separate bridge driver. Patches 11/16 to 16/16 then update all the
>>>>> device tree sources to the new DU and LVDS encoders bindings.
>>>>>
>>>>> I decided to go for live DT patching in patch 08/16 because implementing
>>>>> support for both the legacy and new bindings in the driver would have
>>>>> been very intrusive, and prevented further cleanups. This version relies
>>>>> more heavily on overlays to avoid touching the internals of the OF core
>>>>> compared to v2, even if manual fixes to the device tree are still
>>>>> needed.
>>>>>
>>>>> Compared to v3, this series uses the OF changeset API to update
>>>>> properties instead of accessing the internals of the property structure.
>>>>> This removes the local implementation of functions to look up nodes by
>>>>> path and update properties. In order to do this, I pulled in Pantelis'
>>>>> patch series titled "[PATCH v2 0/5] of: dynamic: Changesets helpers &
>>>>> fixes" at Rob's request, and rebased it while taking two small review
>>>>> comments into account.
>>>>
>>>> Wait a minute!  Why are you putting a patch set to modify core devicetree
>>>> in the middle of a driver series.  Please pull it out to a separate
>>>> series.
>>>
>>> Because Rob asked for the driver-local implementation of the property add
>>> function to be replaced by Pantelis' series. I want to get the LVDS
>>> changes in v4.17 and asked Rob whether I could then take the OF changeset
>>> patches merged through the DRM tree, and he didn't object. If that causes
>>> an issue I'll switch back to the driver-local implementation to get the
>>> driver changes merged, split the OF changeset series out, and then move
>>> to the OF changeset API once merged. Would you prefer that ?
>>
>> You have already created a new version of the R-Car patches without the
>> set of patches that I was objecting to here.  So this is somewhat just
>> an academic comment.
>>
>> As I mentioned in the v6 thread, I am coming back here to clean up loose
>> ends, and explain why I had the reaction I had.  Basically, this is a
>> process issue to me.
>>
>> (1) The patch set from Pantelis is "hidden" in the driver patch series.
>> When viewing collapsed threads (which is my normal mode to avoid getting
>> overwhelmed by the vast volume of email I scan), the Pantelis patch set
>> is totally invisible.  If the R-Car driver patch series had not had me
>> on the to: list, there is a very good chance I would not have noticed
>> it.  Or noticed in a more delayed time frame.  And the same applies to
>> anyone else who might be subscribed to the devicetree mail list.  If
>> the Pantelis patch series was split out into a separate patch set then
>> it would be more visible on the list.
>>
>> (2) There is no good way to indicate in the email subject lines for
>> the Pantelis patches that they are version 3 of the series, since
>> they are also version 4 of the R-Car patch series.  If one reads the
>> patch 0/0 header carefully, and/or the other Pantelis patch comment
>> headers carefully, and then does a little detective work, it is
>> possible to find version 2 of the Pantelis patch series, and thus
>> be able to track the full history.  If just glancing at each
>> individual patch email subject, scanning the patch comment, and
>> spending more time reading the body of the patch, it would be
>> very easy to overlook the existence of the previous versions on
>> the mail list.
>>
>> (2b) Small quibble:  if the Pantelis patches were in a separate series,
>> with v3 in the subject header, then you would have normally put
>> a "changes since v2" section in the patch comment header, which would
>> have been more visible and less cryptic than what you wrote, which was:
>>
>>    ...
>>
>>    Signed-off-by: Pantelis Antoniou <pantelis.antoniou@xxxxxxxxxxxx>
>>    [Fixed memory leak in __of_changeset_add_update_property_copy()]
>>    Signed-off-by: Laurent Pinchart
>> <laurent.pinchart+renesas@xxxxxxxxxxxxxxxx>
>>    ---
>>     drivers/of/dynamic.c | 222 ++++++++++++++++++++++++++++++++++
>>     include/linux/of.h   | 328 +++++++++++++++++++++++++++++++++++++++++++++
>>     2 files changed, 550 insertions(+)
>>
>> My original scan of version 2 and then the email in this series had me
>> questioning whether the two open issues from patch 2 had been addressed
>> (yes, patch 0/16 did explicitly mention that 2 review comments had been
>> taken into account, so you can point at my poor reading comprehension).
> 
> All points dully noted, I'll make sure not to include similar patches in the 
> middle of a driver series should I need to rework a patch series previously 
> posted by someone else.
> 
> I assume your comments also apply to patches to drivers/of/ that are developed 
> as part of a drivers series, not merely pulled in it ?

Yes, as a normal process.  Of course we find cases where following normal process
causes problems, and it can be better to do something different in those cases.


> I have mixed feelings 
> about this, as in many subsystems changes to the core require at least one 
> user. Always splitting core and driver changes in separate series would thus 
> often be impractical.

If at all reasonable, I would still prefer to see two patch series for this
case.  I would expect both series to be upstreamed by a single maintainer
(with the two involved maintainers coordinating this process).  I have not
thought this through in any detail, so I don't know if I'm missing a bunch
of issues.  One issue that I can think of is clearly documenting which
version of the subsystem patch series any given version of the driver patch
series depends upon.  That should be easy to note in patch 0 of the driver
series.  Do you have experience with any other issues in this scenario?

This sounds like a good topic to include in the "Guide to Being A Linux Kernel
Maintainer".


> I doubt we will find a single process that will suit all subsystem 
> maintainers, so maybe this should be a per-subsystem decision ?

Totally agree.  And even for me, I see it as a guideline or normal process,
with exceptions made where it makes sense.


>> (3) This is totally unrelated to your patch series, but I'm leaving a bread
>> crumb trail here.  Rob pointed you at a patch series that was not the most
>> recent version.  When this patch series appears again, there already is
>> a version 3 subset of it, but it is still labeled as being version 2, and
>> also has outstanding un-addressed issues.  ("[PATCH v2 1/2] of: dynamic:
>> Add __of_node_dupv()", 11/04/16)
> 
> Ah, thank you for the information. I was looking for a v3 but missed that one 
> as it was labeled v2.
>