[OE-core] Machine dependencies validation with sstate

[Richard Purdie]

In the process of diving into the issues in bug 1916
(http://bugzilla.yoctoproject.org/show_bug.cgi?id=1916), I've come up
with a good way of detecting machine specific changes to generic
PACKAGE_ARCH packages.

The idea is to have two identical machines, qemux86 and qemux86copy.
These can be setup by:

$ cp meta/conf/machine/qemux86.conf meta/conf/machine/qemux86copy.conf 
[Add SRCREV entry for qemux86copy to meta/recipes-kernel/linux/linux-yocto_3.0.bb]
$ cp -r meta/recipes-core/netbase/netbase-4.47/qemux86 meta/recipes-core/netbase/netbase-4.47/qemux86copy
[Add PACKAGE_ARCH_qemux86copy to meta/recipes-core/netbase/netbase_4.47.bb]
[Add RRECOMMENDS_${PN}_qemux86copy to meta/recipes-multimedia/gstreamer/gstreamer_0.10.35.bb]

You can then generate all of the sigdata files for two machine builds by
running the following commands. The nice thing about these is that you
don't have to run a complete build, it just writes out the data the
build would have generated:

rm tmp/stamps/*/*sigdata*
MACHINE=qemux86 bitbake core-image-sato -S
find tmp/stamps/i586-poky-linux/ -name \*sigdata* | sort > l1
find tmp/stamps/all-poky-linux/ -name \*sigdata* | sort >> l1
MACHINE=qemux86copy bitbake core-image-sato -S
find tmp/stamps/i586-poky-linux/ -name \*sigdata* | sort > l2
find tmp/stamps/all-poky-linux/ -name \*sigdata* | sort >> l2

and then comparing the files l1 and l2, you can see which sigdata files
differ. Running bitbake-diffsigs on the files, e.g.:

bitbake-diffsigs tmp/stamps/all-poky-linux/x11-common-0.1-r44.do_package_write_ipk.sigdata.*

can show what changed and you can then consider whether that is a valid
change and how it might be avoided if necessary. This is how some of the
patches I've just posted were developed.

Comparing qemux86 and qemux86copy as above with core-image-sato now
yields only two differences, in the package_write* tasks of iptables and
gstreamer. This is due to the dependencies these two recipes have on
kernel-module-* packages which in turn depend on linux-yocto which is
machine specific. As yet I can't find a good way to avoid the kernel
dependencies. If we could break debian.bbclass's hold in the
kernel-module case (which never get renamed), that would be one way to
avoid this problem.

I thought people might find this intersting and it was worth documenting
in the list archives if nothing else.

Cheers,

Richard