On 16 March 2017 at 22:11, Kevin Kofler <kevin.kofler@xxxxxxxxx> wrote:
[..]
FYI, this is what Mandriva (back then), now Mageia and OpenMandriva, have
been doing for years (and still do).
https://wiki.mageia.org/en/Overlinking_issues_in_ packaging
> As long as I've done many times full recompilation on the scale of whole
> distribution in the past I know that sometimes is necessary to add small
> linker fix when library X is linked with other libraries A and B and only
> A is used and linked executable is using libX and libB ABI but has no -lB
> in linker options, but majority of those changes have been already merged
> to most of the packages source trees.
Indeed, the folks involved in the Mandriva --as-needed migration reported
that the effort required was similar to the one for the "no more transitive
linking" change
https://fedoraproject.org/wiki/Features/ ChangeInImplicitDSOLinking
in Fedora (since Fedora 13) that you are alluding to.
I've started in PLD those changes when Fedora was in version 5 or 6 :o)
So by this Fedora had less to fix. I remember that on scale those ~6k source packages it was necessary to fix about 200 (maybe 300).
After PLD IIRC it was introduced in Gentoo then OpenMandriva and OpenSuse (or something like that).
I'm not sure how it is now with Debian/Ubuntu but I would not surprised if they started plowing with --as-needed in meantime as well.
Nevertheless after so many whole distro scale iteration now introduction of --as-needed probably will be very easy but still some number of packages may carry incorrect executables linker -l<foo> lists.
I'm thinking only now about yet another kind of pre step before patching ld to make --as-needed default and -fno-as-needed optional.
I think that it would be good to add to koji proper observation deck of those changes.
Simple IMO koji should start collecting informations about at least each package REQUIRES and PROVIDES to track changes as regression tests between each package new releases/revisions.
Tracking those changes would IMO incredibly useful on keeping whole distribution internal entropy under proper control.
Extended version collected data to perform kodi SONAME regression test: not packages dependencies but generate each ELF package binary SONAME list.
In future probably it could be extended to track perl, python and other dependencies not on package layer but with tracking per each .pl/.pm/.py files.
Tracking over koji binary packages REQUIRES may allow track all manually added Requires list changes.
Massive rebuilds logs may be useful to produce whole distro changes statistics.
Probably few next sub ideas could be easy to add as well :)
So this email is kind of call for help to kojii developers :)
I'm not sure how it would with the be best method storing such data on some koji backend services side.
Probably at least additional text file (on top of current three) in current list of files preserved by koji per build request should enough.
Probably SQL backend could be overkill but from my perspective it is implementation detail.
BTW OpenSuSE now uses very dirty patch to reverse ld --as-needed behavior:
$ cat binutils-build-as-needed.diff
Index: ld/ldmain.c
===================================================================
--- ld/ldmain.c.orig 2016-01-27 13:44:14.092983985 +0100
+++ ld/ldmain.c 2016-01-27 13:44:37.389254054 +0100
@@ -286,6 +286,8 @@ main (int argc, char **argv)
#ifdef DEFAULT_FLAG_COMPRESS_DEBUG
link_info.compress_debug = COMPRESS_DEBUG_GABI_ZLIB;
#endif
+ if (getenv ("SUSE_ASNEEDED") && atoi(getenv ("SUSE_ASNEEDED")) > 0)
+ input_flags.add_DT_NEEDED_for_regular = TRUE;
ldfile_add_arch ("");
emulation = get_emulation (argc, argv);
So default --as-need is switched on over SUSE_ASNEEDED env variable.
IMO something like this could be acceptable only as JFDI/JFDIN solution.
Definitely modifications of rpm %configure and cmake related macros will not cover everything and ld patch is only IMO logical/proper solution.
Such proper solution should simple change declared input_flags.add_DT_NEEDED_for_regular to TRUE and add printing on stdout linker warning about not use exact SONAME or as alternative that some exact -l<foo> was not used/ignored because now using --as-needed does not shows what exactly is dropped.
IMO patch like above may have high chance to be accepted by binutils source tree maintainers and should be pushed ASAP to binutils source tree.
When koji and ld patch will be integrated in binutils as bcond disabled by default such bcond on time +0T could be removed and yet another massive rebuild pushed to build systems with binutils first in the queue.
Keeping such patch before final push will allow more developers test new ld behavior by something like "rpmbuild --rebuild --with rev-as-needed binutils*src.rpm" and force reinstall binutils package on own system.
So this email as well is kind of call to have green light from core Fedora developers/maintainers committee that as it is described here it looks acceptable to start working on details.
(Simple I don't want to waste other people and my time if such proposal will be refused .. but IMO it is really worth because it is kind of game changer).
I'm not sure am I introducing such proposal correctly.
If not please let me know (priv) what is TheRightWay(tm).
Comments?
kloczek
--
Tomasz Kłoczko | LinkedIn: http://lnkd.in/FXPWxH
_______________________________________________ devel mailing list -- devel@xxxxxxxxxxxxxxxxxxxxxxx To unsubscribe send an email to devel-leave@xxxxxxxxxxxxxxxxxxxxxxx
