[OE-core] Kernel: Builds use old version of defconfig

[Khem Raj]

> On Nov 28, 2016, at 11:58 PM, Mike Looijmans <mike.looijmans at topic.nl> wrote:
> 
> On 29-11-16 03:03, Khem Raj wrote:
>> 
>>> On Nov 24, 2016, at 5:55 AM, Bruce Ashfield <bruce.ashfield at gmail.com
>>> <mailto:bruce.ashfield at gmail.com>> wrote:
>>> 
>>> 
>>> 
>>> On Thu, Nov 24, 2016 at 5:32 AM, Mike Looijmans <mike.looijmans at topic.nl
>>> <mailto:mike.looijmans at topic.nl>> wrote:
>>> 
>>>    On 24-11-16 11:10, Mike Looijmans wrote:
>>> 
>>>        I'm currently experiencing a problem with "defconfig" files and the
>>>        kernel.
>>> 
>>>        In short, when I make a change to the "defconfig" file, the kernel
>>>        is rebuilt
>>>        (which is correct), but the resulting kernel has been built using
>>>        the old
>>>        defconfig from a previous build, instead of the new one.
>>> 
>>>        The kernel recipe just contains "file://defconfig" in its SRC_URI. The
>>>        defconfig file is in the project's overlay.
>>> 
>>>        For example, I have a kernel with "CONFIG_DEVMEM" disabled:
>>> 
>>>        # gunzip < /proc/config.gz | grep DEVMEM
>>>        # CONFIG_DEVMEM is not set
>>> 
>>>        Now, I change the defconfig to contain CONFIG_DEVMEM=y and build the
>>>        image.
>>>        The result:
>>> 
>>>        # gunzip < /proc/config.gz | grep DEVMEM
>>>        # CONFIG_DEVMEM is not set
>>> 
>>>        So the change did not make it into the actual kernel, even though
>>>        the kernel
>>>        was rebuild as a result of the change.
>>> 
>>>        I run "bitbake -c cleansstate virtual/kernel" and build the image again:
>>> 
>>>        # gunzip < /proc/config.gz | grep DEVMEM
>>>        CONFIG_DEVMEM=y
>>> 
>>>        After cleaning, the result is correct and the new defconfig is active.
>>> 
>>>        I'm trying to figure out how this can happen, any help is welcome...
>>> 
>>> 
>>>    What seems to be the problem is this code in kernel.bbclass:
>>> 
>>>            # Copy defconfig to .config if .config does not exist. This allows
>>>            # recipes to manage the .config themselves in
>>>    do_configure_prepend().
>>>            if [ -f "${WORKDIR}/defconfig" ] && [ ! -f "${B}/.config" ]; then
>>>                    cp "${WORKDIR}/defconfig" "${B}/.config"
>>>            fi
>>> 
>>>    This keeps any existing ".config" file if it happens to still be in the
>>>    $B path, which is the case if you're rebuilding a kernel.
>>> 
>>>    I see two possible ways to fix this.
>>> 
>>>    1) During "cleanup" also remove the .config file in the build dir.
>>>    However, the build dir is probably kept alive for a reason? I also can't
>>>    figure out how that "cleanup" is being done.
>>> 
>>> 
>>>    2) Remove the second part of the "if" statement, so it becomes:
>>> 
>>>            # Copy defconfig to .config if "defconfig" exists. This allows
>>>            # recipes to manage the .config themselves in
>>>    do_configure_prepend().
>>>            if [ -f "${WORKDIR}/defconfig" ]; then
>>>                    cp "${WORKDIR}/defconfig" "${B}/.config"
>>>            fi
>>> 
>>>    I've tested that, and it solves my problem. However, it will probably
>>>    break other people's config mangling?
>>> 
>>> 
>>> Yep, in particular all the fragment processing which has the capability of
>>> starting
>>> with a defconfig and then apply fragments from any number of other places. When
>>> that task completes, a full .config is in ${B}.  If that statement comes
>>> along and
>>> clobbers the .config …
>> 
>> so you either assume that .config is valid once generated or you dont. When a
>> configure task
>> is triggered it should recreate .config everytime.
> 
> The problem seems to be that the class "do_configure" does things that should happen before and after things that the recipe would want to change.
> 
> Copying defconfig or whatever means to create a .config should be first.
> 
> Next, the specific kernel recipe would want to mangle that configuration to suit its needs, like applying fragments and such.
> 
> Then the makeoldconfig (or whatever) task should run.



> 
> 
> The current system assumes that the kernel recipe creates a do_configure_prepend to do the mangling, which is rather counterintuitive, one would expect to "append" extra actions.
> 
> A structured approach would be to split the do_configure into two parts that should run in sequence, and then kernel recipes can inject their actions by appending to them as they see fit. The first task would create the .config file by (forcibly) copying any defconfig or starting point. The second task would call the kernel's make script to futher process it.
> 

it might just be enough to expect kernel.bbclass to error out if a given task is not implemented by inheriting recipes much like pure virtual functions in C++
and get rid of _append/_prepend sequencing expectations.


> But this too would break existing recipes.
> 
> 
> 
>>> 
>>> I'm actually working in the 2.3 release cycle to make the fragment processing
>>> be available to all kernels, which will likely solve this problem .. but we
>>> can't
>>> wait for that.
>>> 
>>> So I'm hoping that there's a way to make the behaviour cover both use cases.
>>> 
>>> Maybe someone with more bitbake knowledge can point out a way that can
>>> detect if the task is being run due to a change in the task signature.
>>> 
>>> Since if you've modified the defconfig, the task is being re-run for that change
>>> and at that point we could safely remove the .config (versus forcing it on the
>>> clean step).
> 
> 
> 
> 
> Kind regards,
> 
> Mike Looijmans
> System Expert
> 
> TOPIC Products
> Materiaalweg 4, NL-5681 RJ Best
> Postbus 440, NL-5680 AK Best
> Telefoon: +31 (0) 499 33 69 79
> E-mail: mike.looijmans at topicproducts.com
> Website: www.topicproducts.com
> 
> Please consider the environment before printing this e-mail
> 
> 
> 
> 
>