[poky] kernel.release being regenerated

[Bruce Ashfield]

On 15-02-11 04:10 AM, Richard Purdie wrote:
> On Wed, 2015-02-11 at 09:27 +0100, Holger Hans Peter Freyther wrote:
>> On Tue, Feb 10, 2015 at 10:51:26PM +0000, Richard Purdie wrote:
>>> Hi Holger,
>>
>> Hi Richard,
>>
>>> The fact the kernel.release is being recreated suggests something in the
>>> configuration is changing (different environment or commandline
>>> options?) or that there is a problem in your kernel to do with the
>>> Makefiles and the dependencies of the kernel.release file.
>>
>> This issue should be seen by everyone but the race is pretty small.
>> do_shared_workdir must run shortly after compilemodules has started
>> and removed the file but didn't create a new one. This is the kind
>> of race I am more likely to hit than the average population. The
>> "configuration is changing" part looks like a red herring to me.
>>  From looking at the Makefile I see a FORCE as dependency of the
>> kernel.release rule. With my 3.10er kernel (and it doesn't look
>> like we have patched the Makefile) the dependency chain is like
>> this:
>>
>>   init -> prepare -> prepare0 .. -> prepare3 -> kernel.release -> FORCE
>>
>> This means that on any target being executed the kernel.release
>> will be re-created. I think the only stable way would be to have
>> do_compile copy the kernel.release to another place and have the
>> copying task pick this file from another place.
>
> This does sound like a flaw in how we're copying the file if it is
> regenerated on each execution of make within the kernel build.
>
> What puzzles me is that the use of kernel.release isn't new, the
> previous do_install code used to do this and technically would have
> raced with do_compilemodules too. So was it just bad luck you started
> hitting this now?

It looks like bad luck that this is popping up now.

>
> (http://git.yoctoproject.org/cgit.cgi/poky/tree/meta/classes/kernel.bbclass?id=92725ad46f4d331bea6a2fa65964158d78a7add8)
>
> As you say, if this really is getting regenerated as you describe, we'll
> have to store a safe copy in the tree.

What about changing the dependencies of the known tasks ? i.e. making
sure that the kernel modules don't start building until we have the
shared workdir populated ? Otherwise, what does that leave us, staging
a copy at the end of kernel_do_compile and grabbing it from there in the
shared workdir population step ?

Bruce

>
> Cheers,
>
> Richard
>