[OE-core] Can we drop eglibc-utils from LIBC_DEPENDENCIES?

Sat Dec 17 01:29:27 UTC 2011

On Fri, 2011-12-16 at 15:34 -0800, Darren Hart wrote:
> 
> On 12/16/2011 03:20 PM, Darren Hart wrote:
> > 
> > 
> > On 12/16/2011 01:07 PM, Koen Kooi wrote:
> >>
> >> Op 16 dec. 2011, om 19:30 heeft Darren Hart het volgende geschreven:
> >>
> >>> I'm working on a minimal distro definition, and found that eglibc-utils
> >>> pulls in bash (needed for tzconfig and xtrace apparently)
> >>
> >> My first thought is: fix the bashisms in those scripts, I bet ubuntu/fedora/arch/gentoo have patches for that,
> > 
> > 
> > Agreed, this would be a good thing to do. However, I still shouldn't
> > need to include this in a "tiny" distribution.
> > 
> > 
> >>> which pulls in
> >>> gettext, which requires wchar support. I'd like to remove eglibc-utils
> >>> from my distro definition. I could override the default I suspect, but I
> >>> wonder if eglibc-utils should be made an optional package that distro
> >>> definitions, images, or users should specifically add if needed?
> >>>
> >>> The relevant bit of code appears to be:
> >>>
> >>> meta/conf/distro/include/tclibc-eglibc.inc
> >>>
> >>> LIBC_DEPENDENCIES = "libsegfault \
> >>> 		     eglibc \
> >>>    		     eglibc-dbg \
> >>>    		     eglibc-dev \
> >>>    		     eglibc-utils \
> >>>    		     eglibc-thread-db \
> >>>    		     eglibc-localedata-i18n \
> >>>    		     eglibc-gconv-ibm850 \
> >>>    		     eglibc-gconv-cp1252 \
> >>>    		     eglibc-gconv-iso8859-1 \
> >>>    		     eglibc-gconv-iso8859-15 \
> >>>    		     locale-base-en-us \
> >>>    		     locale-base-en-gb "
> >>>
> >>> eglibc-dbg and eglibc-dev also seem like they could be made optional.
> >>>
> >>> Thoughts? Would anyone object to me removing at least eglibc-utils from
> >>> LIBC_DEPENDENCIES?
> >>
> >> I did a little digging:
> >>
> >> koen at dominion:/OE/tentacle/sources/openembedded-core$ git grep LIBC_DEPENDENCIES
> >> meta/conf/distro/include/tclibc-eglibc.inc:LIBC_DEPENDENCIES = "libsegfault \
> >> meta/conf/distro/include/tclibc-uclibc.inc:LIBC_DEPENDENCIES = "\
> >> meta/recipes-core/tasks/task-core-nfs.bb:GLIBC_DEPENDENCIES = "glibc-utils"
> >> meta/recipes-core/tasks/task-core-nfs.bb:RRECOMMENDS_task-core-nfs-server_append_libc-glibc = " ${GLIBC_DEPENDENCIES}"
> >> meta/recipes-core/tasks/task-core-standalone-sdk-target.bb:    ${LIBC_DEPENDENCIES} \
> >>
> >> So it's only used for debug and/or SDK uses. I am going to argue that if you're going to support debug and SDK you're not minimal anymore and can live with bash/gettext/etc.
> > 
> > Well, nfs isn't SDK only, there are valid deployment uses for that. But
> > otherwise, agreed.
> > 
> >>
> >> Since I was bored I dug up an OE-classic:
> >>
> >> koen at dominion:/OE/org.openembedded.dev$ git blame  recipes/tasks/task-sdk-bare.bb
> >> 749310c7 packages/tasks/task-sdk-bare.bb (Koen Kooi 2008-07-07 21:24:38 +0000  1) DESCRIPTION = "Packages for a standalone  SDK or external toolchain"
> >> [..]
> >> 9bff47f7 packages/tasks/task-sdk-bare.bb (Tom Rini  2008-11-26 13:16:21 -0500  8) GLIBC_PKGS = "\
> >> 749310c7 packages/tasks/task-sdk-bare.bb (Koen Kooi 2008-07-07 21:24:38 +0000  9)     glibc \
> >> 749310c7 packages/tasks/task-sdk-bare.bb (Koen Kooi 2008-07-07 21:24:38 +0000 10)     glibc-dbg \
> >> 86fa8521 packages/tasks/task-sdk-bare.bb (Tom Rini  2009-02-04 02:07:47 -0500 11)     virtual-libc-dev \
> >> 749310c7 packages/tasks/task-sdk-bare.bb (Koen Kooi 2008-07-07 21:24:38 +0000 12)     glibc-utils \
> >> 749310c7 packages/tasks/task-sdk-bare.bb (Koen Kooi 2008-07-07 21:24:38 +0000 13)     libsegfault \
> >> 749310c7 packages/tasks/task-sdk-bare.bb (Koen Kooi 2008-07-07 21:24:38 +0000 14)     glibc-thread-db \
> >> f18a05e2 recipes/tasks/task-sdk-bare.bb  (Tom Rini  2010-02-09 16:43:45 -0700 15) "
> >> 9bff47f7 packages/tasks/task-sdk-bare.bb (Tom Rini  2008-11-26 13:16:21 -0500 16) 
> >> edd3a1de recipes/tasks/task-sdk-bare.bb  (Tom Rini  2011-01-18 17:56:52 -0700 17) LIBC_PKGS_libc-glibc = "${GLIBC_PKGS}"
> >> edd3a1de recipes/tasks/task-sdk-bare.bb  (Tom Rini  2011-01-18 17:56:52 -0700 18) LIBC_PKGS_libc-uclibc = "uclibc uclibc-dev uclibc-thread-db"
> > 
> > Was this list used in the same way as LIBC_DEPENDENCIES above?
> > 
> >>
> >> So a few years ago that list of packages was only meant for SDK usage.
> >>
> >> If you meant GLIBC_DEPENDENCIES (note the extra 'G'), then you need
> >> to
> >> check if they are still needed for NFS operation. If so I am going to
> >> argue that the dependencies should move to the recipes in question
> >> instead of hiding in the task.
> > 
> > Right, that makes sense.
> > 
> >> If it's just a convenience package go
> >> ahead and remove it, people wanting it can create a new task :)
> > 
> > Agreed as well.
> > 
> > I ran into an interesting issue. If I remove eglibc-utils from
> > LIBC_DEPENDENCIES, it still seems to be getting pulled in, as do bash
> > and gettext. Still digging to sort out why...
> 
> It would appear that removing it from LIBC_DEPENDENCIES prevents it from
> being installed, but it is still built and, worse, so are all of it's
> RDEPENDS, which pull in bash and gettext and then break on a small libc
> with no widechar support.
> 
> So, is it correct behavior to build RDEPENDS for packages that will not
> be installed?
> 
> If so, I gather my fix is to remove eglibc-utils from the packages
> generated by the eglibc recipe when building with my tiny distro?

What I suspect you're seeing is something like a recipe which does:

PACKAGES = "A B"
RDEPENDS_A = "1 2 3"
RDEPENDS_B = "2 3 4"

and that you're using A but not B. Since the build system needs to build
A and B at the same time as they're part of the same recipe, it will
build 1-4. It won't necessarily install them.

Usually, if this is causing conflict we'd split the recipe up.

Cheers,

Richard