Re: [poky] [PATCH 04/13] module: build hostprogs for each module

[Darren Hart <dvhart@linux.intel.com>]

On 03/14/2011 11:10 AM, Koen Kooi wrote:
>
> Op 7 mrt 2011, om 17:30 heeft Darren Hart het volgende geschreven:
>
>> On 03/04/2011 03:20 PM, Darren Hart wrote:
>>> On 03/04/2011 02:30 PM, Koen Kooi wrote:
>>>>
>>>> Op 4 mrt 2011, om 23:28 heeft Koen Kooi het volgende geschreven:
>>>>
>>>>>
>>>>> Op 4 mrt 2011, om 22:44 heeft Darren Hart het volgende geschreven:
>>>>>
>>>>>> On 03/04/2011 12:04 PM, Koen Kooi wrote:
>>>>>>>
>>>>>>> Op 4 mrt 2011, om 20:42 heeft Saul Wold het volgende geschreven:
>>>>>>>
>>>>>>>> From: Darren Hart<dvhart@linux.intel.com>
>>>>>>>>
>>>>>>>> This fixes [BUGID #241]
>>>>>>>>
>>>>>>>> The kernel hostprogs are built for the host architecture. They
>>>>>>>> should not be
>>>>>>>> deployed to the target, and they should not be included in an
>>>>>>>> sstate package
>>>>>>>> which might get reused on a host of a different architecture.
>>>>>>>>
>>>>>>>> As we don't build many out-of-tree modules, this patch takes the
>>>>>>>> approach of
>>>>>>>> building the hostprogs as part of the module compile process with a
>>>>>>>> do_compile_prepend() routine in module.bbclass.
>>>>>>>>
>>>>>>>> We don't have to clean the hostprogs as modules depend on the
>>>>>>>> kernel being
>>>>>>>> populate_staging, so its done with the staging directory by the
>>>>>>>> time we run.
>>>>>>>>
>>>>>>>> Signed-off-by: Darren Hart<dvhart@linux.intel.com>
>>>>>>>> CC: Gary Thomas<gary@mlbassoc.com>
>>>>>>>> ---
>>>>>>>> meta/classes/module.bbclass | 12 +++++++++++-
>>>>>>>> 1 files changed, 11 insertions(+), 1 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/meta/classes/module.bbclass
>>>>>>>> b/meta/classes/module.bbclass
>>>>>>>> index d16d462..bbceaf7 100644
>>>>>>>> --- a/meta/classes/module.bbclass
>>>>>>>> +++ b/meta/classes/module.bbclass
>>>>>>>> @@ -3,6 +3,13 @@ DEPENDS += "virtual/kernel"
>>>>>>>>
>>>>>>>> inherit module-base
>>>>>>>>
>>>>>>>> +# Ensure the hostprogs are available for module compilation
>>>>>>>> +module_do_compile_prepend() {
>>>>>>>> + unset CFLAGS CPPFLAGS CXXFLAGS LDFLAGS
>>>>>>>> + oe_runmake CC="${KERNEL_CC}" LD="${KERNEL_LD}" AR="${KERNEL_AR}" \
>>>>>>>> + -C ${STAGING_KERNEL_DIR} scripts
>>>>>>>> +}
>>>>>>>> +
>>>>>>>> module_do_compile() {
>>>>>>>> unset CFLAGS CPPFLAGS CXXFLAGS LDFLAGS
>>>>>>>> oe_runmake KERNEL_PATH=${STAGING_KERNEL_DIR} \
>>>>>>>> @@ -15,7 +22,10 @@ module_do_compile() {
>>>>>>>
>>>>>>> Why not put it in do_compile itself?
>>>>>>
>>>>>> My rationale was that a module recipe may override do_compile for
>>>>>> one reason or another, but they will still need the hostprogs which
>>>>>> we stripped in kernel.bbclass. This way, even if they override
>>>>>> do_compile, the _prepend will still get run and prepare the
>>>>>> hostprogs.
>>>>>
>>>>> Are you sure that actually works? AIUI bitbake will fold the compile
>>>>> and prepend into one when you try to override it.
>>>>
>>>> What I actually meant to say is: If you need to prepend it to
>>>> do_compile, even when someone will override it, use addtask to insert
>>>> it between configure and compile. That makes the intention a lot
>>>> clearer and isn't depending on faulty bitbake behaviour.
>>>
>>> You could be right, I'm not terribly familiar with how the "inheritance"
>>> model works. It seems to me the "addtask" method would work regardless
>>> as you suggest, and I have no objection to taking that route.
>>>
>>> Richard, do you have a preference? Can you clear up the inheritance
>>> question?
>>>
>>
>> Had a chat with Richard. His suggestion was to use a new function, ie do_hostprogs, which do_compile() calls. If a recipe overrides do_compile() it can call do_hostprogs() or manage it on its own. I'll do some testing and send out a follow-on patch.
>
> Any news on this? I just hit the issue:
>
> |   CC [M]  /OE/tentacle/build/tmp-angstrom_2010_x/work/beagleboard-angstrom-linux-gnueabi/ti-dsplink-1_1_65_00_03-r99i/dsplink_linux_1_65_00_03/dsplink/gpp/src/../../gpp/src/arch/CFG_map.o
> | /bin/sh: scripts/basic/fixdep: No such file or directory

Yes, I have a working solution, but have been trying to polish it up and 
getting pulled away on other bugs. I'll clean it up and push it out by 
tonight, although I expect a few review cycles will be needed before we 
are all happy with it.

The tricky part is balancing the complexity of building the configured 
source tree with the size it takes up in sysroots. It can be done in as 
little as 80MB, but the easiest way (and the easiest to maintain going 
forward) is 500MB. More on that tonight.

-- 
Darren Hart
Intel Open Source Technology Center
Yocto Project - Linux Kernel