On 5/24/13 3:52 AM, Paul Eggleton wrote:
On Thursday 23 May 2013 09:40:51 Mark Hatle wrote:
On 5/23/13 3:01 AM, Ming Liu wrote:
libpam might miss ABI specific dependencies for pam-plugins-*, for RPM
uses
generic names to check the packages depending on it and doesn't consider
the arch, which will lead to packaging issues in multilib build.
pam_plugin_hook is added because the plugin packages are dynamically
generated, so we need to manually process multilib names by add baselib to
RPROVIDES/RDEPENDS as ABI specific tag.
[YOCTO #4532]
[ CQID: WIND00416824 ]
Signed-off-by: Ming Liu <[email protected]>
I worked with Ming Liu on this particular issue. You may wonder why this is
necessary let me attempt to explain the underlying causes.
In deb/ipk on a multilib package, the package name has specific multilib
references in it. I.e. the alternative libraries start with something like
lib32-... This was done primarily because deb/ipk do not allow two packages
with the same name (but different architectures) to be installed at the
same time. So the name has to be unique.
In RPM however, the names of the packages and matches with the architectures
and if they are not the same we can do these multilib installs. This
matches the behavior of other RPM based distributions and in many ways the
tools people are used to working with RPM. For the most part this works
fine in multilib configurations because additional per-file dependencies
are added that capture the shared library dependencies with ABI specific
information. This unfortunately fails in a few cases where plugins are
dynamically loaded via dlopen -- such as libpam.
One possible fix is simply to follow the deb/ipk package naming, but this
causes a design advantage of rpm. When a package has a dependency on
'bash', we really don't care what bash is installed, only that -a- bash is
installed. In the deb/ipk case, the lib32- packages would end up with a
lib32-bash dependency and you could potentially end up with two 'bash'
packages being installed.
So the fix I recommended for the issue was to add the baselib path to the
internal dependencies. Since we know that the libpam installed in 'lib'
needs the modules that were compiled to also work with the 'lib' version of
libpam. While the libpam in 'lib64' need the modules to work with the
'lib64' version of the plugins.
Existing dependencies are preserved so there is no impact in the ipk/deb
case, the RPM case is resolved as the additional dependency information is
now present for the package manager to select the package we really want.
If anyone else has a suggestion for an alternative fix, we're interested --
but this is the best answer we could come up with. (If any of the above
should be added to the commit message, the YP bug, or documentation, please
let me know and I'll make sure it gets added.)
This is the same problem as bug 4408:
https://bugzilla.yoctoproject.org/show_bug.cgi?id=4408
It's a bit nasty to have to solve this on a per-recipe basis; we need some
kind of generic solution. At the moment I'm not sure what that would be
however.
Yes this would be the same issue. I'm really not sure how to do this
automatically as it requires knowledge that the recipe has plugins, and that the
plugs are ABI/lib specific.
I'm definitely open to suggestions.
--Mark
Cheers,
Paul
_______________________________________________
Openembedded-core mailing list
[email protected]
http://lists.openembedded.org/mailman/listinfo/openembedded-core
