[PATCH rtems 00/12] bsp/imxrt: Update SDK and prepare for new variant

[Chris Johns]

On 15/5/2023 6:13 pm, oss at c-mauderer.de wrote:
> Hello Chris,
> 
> Am 15.05.23 um 05:18 schrieb Chris Johns:
>>
>>
>> On 6/5/2023 2:02 am, Gedare Bloom wrote:
>>> On Fri, May 5, 2023 at 9:02 AM <oss at c-mauderer.de> wrote:
>>>>
>>>> Hello Gedare,
>>>>
>>>> thanks for taking a look at the patch set.
>>>>
>>>> Am 05.05.23 um 15:56 schrieb Gedare Bloom:
>>>>> On Thu, May 4, 2023 at 9:01 AM Christian Mauderer
>>>>> <christian.mauderer at embedded-brains.de> wrote:
>>>>>>
>>>>>> Hello,
>>>>>>
>>>>>> this patch set for the arm/imxrt BSP family updates the SDK files to the
>>>>>> latest version of the mcux-sdk from NXP and prepares the BSP for further
>>>>>> chip variants. I plan to add a BSP that uses the IMXRT1166 soon.
>>>>>>
>>>>>> As a base for the mcux-sdk files, I now use the NXP git repository
>>>>>> instead of zip files that can be downloaded from NXP. I kept the exact
>>>>>> file system structure to make future updates simpler.
>>>>>>
>>>>>> To import the files, I used a script. It is not a clean script and it
>>>>>> was only tested on a Linux machine. Despite that, I added that script to
>>>>>> the BSP directory in case someone else ever wants to update the
>>>>>> mcux-sdk. Updating the SDK is also possible without the script. It's
>>>>>> just a lot more manual work. So if we don't want a script in that state
>>>>>> in the repository, I can also just keep it on a private branch.
>>>>>>
>>>>> I would recommend that you document the import/update process for the
>>>>> SDK in this BSP (or family) documentation. You could provide a stable
>>>>> external link to the script there instead of including it in the
>>>>> rtems.git tree. I'm not convinced that including scripts etc in the
>>>>> same tree as the source they manipulate is a great practice. It seems
>>>>> to violate separation of concerns.
>>>>
>>>> Disadvantage is that the script (and documentation) is a bit less simple
>>>> to find. But that's not a problem for me (I know where to search) so I
>>>> will adapt that.
>>>>
>>>> De we have an official preferred stable location? Otherwise, I think
>>>> either a small repo in my https://git.rtems.org/christianm or a repo on
>>>> github should be a good place (most likely the later one).
>>>>
>>> This should be fine. ftp.rtems.org should also be usable (in theory)
>>>
>>>>>
>>>>>> The patches that import the new SDK files (patch 0002) and remove the
>>>>>> old ones (patch 0004) are too big for the list. I'll only send the
>>>>>> summary. You can find the full patches here:
>>>>>>
>>>>>>     0002:
>>>>>> https://gitlab.com/c-mauderer/rtems/-/commit/2a871672767a95598e5af42373bfebd3eb9440d3
>>>>>>     0004:
>>>>>> https://gitlab.com/c-mauderer/rtems/-/commit/2a3e104fa808d7f34a1930344d7b39d11cf39f3d
>>>>>>
>>>>>> The complete patch set is on this branch:
>>>>>>
>>>>>>     https://gitlab.com/c-mauderer/rtems/-/commits/cm/20230504_imxrt/
>>>>>>
>>>>>> At the moment, I import the support files for all currently available
>>>>>> i.MXRT* variants. The headers for the CPU registers are really big (a
>>>>>> few megabytes per header) which makes the complete source tree of the
>>>>>> mcux-sdk bigger than 90MB. If preferred, I can remove most variants and
>>>>>> only keep the ones that are currently used or will be used soon
>>>>>> (IMXRT1052 and IMXRT1166) to reduce the size.
>>>>>>
>>>>> Thanks, I think this is fine. The increase in the number of HAL/SDK
>>>>> that we have to import is starting to concern me in general. I wonder
>>>>> if there is a better way to manage these external sources.
>>>>>
>>>>
>>>> I think I have a language problem here: Is it fine that I push 96MB or
>>>> is it fine that I reduce the size of the patch before pushing?
>>>>
>>> My apologies, I missed this also. If it's not that much more work, it
>>> might be best to remove the unused variants. We don't like to carry
>>> around dead code in the repo as another general rule.
>>>
>>>> In this case I only updated the existing HAL with more chips, but I
>>>> agree that having a better solution to manage HALs than the
>>>> clone-and-own approach would be great.
>>>>
>>>> One possible solutions could be to use subrepos. We can clone the lib to
>>>> the git.rtems.org, add necessary patches and add that as a subrepo. If
>>>> an update is necessary, patches can just be rebased on the latest
>>>> version of the upstream lib. Another one could be something like the
>>>> west tool that is used in Zephyr. Maybe can also make it simpler to
>>>> include libs with vendor specific licenses.
>>>>
>>>> But approaches like that will have a big impact on how to handle a lot
>>>> of tasks in RTEMS. For example you would have to init submodules before
>>>> building BSPs. The submodules have to be handled during release. And a
>>>> lot more. So we shouldn't discuss that as a side note to a patch set but
>>>> rather as a new mail thread.
>>>>
>>> Agreed.
>>
>> Do these HAL systems provide stable API interfaces? If so is all RTEMS needs to
>> interface to is the API headers?
>>
>> Chris
>>
> 
> I think that depends a lot on the vendor. In my experience, the better HALs are
> mostly stable. For example, I don't think that the ST HAL changed the API in the
> versions that I know. And during the update of the NXP SDK I didn't note a
> breaking API change either. Only addition of new functions.

That is good news. OAR has been adding Xilinx bare metal BSP drivers to the
Microblaze and that code in Vitis is for all Xilinx devices so I am not sure how
that will workout long term. We are no position to make any hard and fast rules
because of the differences so this conversation is not an attempt to do that
rather a prompt for us explore and discussion how we sort this out.

> The HAL provides basic drivers. For example in the i.MXRT BSP I use it for stuff
> like setting up clocks or adding drivers like serial, I2C or SPI. I expect that
> at least some drivers are necessary to link every program. So without the HAL,
> the test suite wouldn't build. As far as I know that would be something new in
> RTEMS that the tests can't be built without external dependencies.

Yes. It is tricky because we need code to start RTEMS however users could
benefit from other pieces code and drivers in a HAL for applications the RTEMS
kernel does not need to start and run a test. Splitting the HAL type code into
pieces does not offer a good long term solution either.

> I don't have a problem with a change like that, but it is something that would
> have a quite big influence on workflows and how we plan to handle similar cases
> in the future.

Yes it does. I see issues currentlt as the absorption of lot and lot of code
like we are starting to do. There is also long term support of systems using it.

> Like said to Gedare: If we want to discuss possible solutions in detail: Please
> let's start a separate thread that doesn't target NXP HAL only but HALs in
> general (ST HAL, ARM CMSIS, NXP HAL, ...). Should I write a mail starting a
> discussion?

I agree. This change is not being singled out. At some point we were going to
tip over an edge and hit an issue.

> I don't expect that we find a solution for that before the next release.

I agree.

> Therefore, I still would like to apply the v2 of the patch set, which is just an
> update of the currently existing HAL in the BSP. Of course if someone suggests a
> better solution that can be included in the release, I'm happy to update the
> patch set to that solution.

Yes. I think you understand my response was not about this change, it was a
reaction to the growing problem we have. I think it is a good one to solve and
it could be a focus for 7 and then we sunset the transition in a later release.
I see Xilinx and the x86 with EFI as other candidates.

Chris