[OE-core] [oe] kernel.bbclass: Fix do_shared_workdir task ordering

[Jens Rehsack]

> Am 14.10.2015 um 21:48 schrieb Bruce Ashfield <bruce.ashfield at gmail.com>:
> 
> On Wed, Oct 14, 2015 at 3:30 PM, S. Lockwood-Childs <sjl at vctlabs.com> wrote:
>> http://patchwork.openembedded.org/patch/99875/
>> 
>> Apparently this patch is still not in master, and I just ran across the
>> problem with an externally built module (omaplfb from omap3-sgx-modules)
>> in meta-ti layer.
>> 
>> What's the plan for getting a correct Module.symver into shared_workdir
>> for external modules to build against? Above patch, or does someone have
>> an even better idea?
> 
> Richard and I sync'd on this while in Dublin @ ELCe, and the changes
> aren't missing
> from master by mistake .. but more because we are still working to come up with
> a comprehensive solution (tracked in bugzilla).
> 
> The solution is pretty much what I described before, we are balancing
> applications
> and tasks that do not need kernel modules to be built, versus external modules
> that depend on symbols from other modules. The devil is in the
> details, and getting
> a non-racy, task locked solution that allows the recipe writer to
> explicitly decide
> whether they need modules built or not .. attempts at detecting the
> need, or forcing
> a one size fits all solution have all lead to dead ends.
> 
> Since we are close to a release point, I'm still working on this out
> of the tree, and
> will propose some changes when the tree looks stable.
> 
> For now, you can carry the patch locally, or you can append to the kernel module
> compilation task and do a second copy of the symvers file to the share
> directory.
> 
> i.e. a variant of this:
> http://patchwork.openembedded.org/patch/94891/, done in a
> bbappend versus the class.
> 
> Cheers,
> 
> Bruce

This is kind of insane to try a fix duplicating a job in a probably wrong way
(even a tiny) because of performance issue.

Any 3rd party kernel module depends on do_shared_workdir - so do_shared_workdir
must wait for do_compile_kernelmodules - that's it by design (build is done
relying on dependencies ordered in a directed, noncyclical graph.

Since compile_kernelmodules is between compile and strip, I vote for

$ git diff
diff --git a/meta/classes/kernel.bbclass b/meta/classes/kernel.bbclass
index 5e8b6cf..49d7561 100644
--- a/meta/classes/kernel.bbclass
+++ b/meta/classes/kernel.bbclass
@@ -253,7 +253,7 @@ kernel_do_install() {
 }
 do_install[prefuncs] += "package_get_auto_pr"

-addtask shared_workdir after do_compile before do_compile_kernelmodules
+addtask shared_workdir after do_compile_kernelmodules before do_strip
 addtask shared_workdir_setscene

 do_shared_workdir_setscene () {

But that's surely kind of smell, whether before do_strip and do_install is
preferred. Mandatory is, that do_shared_workdir must not be processed before
do_compile_kernelmodules finishes.

There is no way to avoid it, and if those 5 seconds slow down your build,
there is probably another thing which should be fixed.

Cheers
-- 
Jens Rehsack - rehsack at gmail.com