On 2017-01-19 01:53, Cyril Brulebois wrote: > Cyril Brulebois <kibi@debian.org> (2017-01-19): > > Summing up things from IRC: > > - in an uptodate sid chroot (both development version and minimal one > > with daily-build script): no DNS issues with the generated mini.iso > > (amd64, tested within stable's kvm on amd64). My image was also > > tested successfully by Steve, so not a setup issue. > > > > - I can reproduce the issue with dailies from 2017-01-18, and from > > 2017-01-19 (I picked it in advance during its build on barriere). > > > > - I can reproduce the issue in the above mentioned sid chroot if > > I downgrade these packages to testing's version (-9 → -8): > > sudo apt-get install libc6:amd64=2.24-8 libc6-dev:amd64=2.24-8 libc-dev-bin=2.24-8 > > > > - The older set of libc* packages is installed in barriere's current > > amd64 sid chroot, too. > > > > - Steve could only produce broken images when building locally, until > > he upgraded his libc* packages as well. > > And since Steve was wondering how we could release something for Stretch > RC 1 with that… I've just confirmed that using -8 .deb with a -8 .udeb > (reversioned as -9+hack so that its being dropped under localudebs makes > it take precedence over sid's .udeb) leads to an image that's working > fine. And AFAICT that's the combination we had at the time. > > I suspect the implementation change through the following patch in -9: > glibc-2.24/debian/patches/any/cvs-resolv-internal-qtype.diff Indeed, I haven't done any test, but looking at the code, it shows that the changes assume that libresolv.so.2 and libnss_dns.so.2 stay in sync. > plus 283e7a294275a7da53258600deaaafbbec6b96c1 in debian-installer.git is > what is triggering the issue? (Explaining why host's libc .deb have an > impact on the built images.) Indeed this has been done to avoid crashes when libc6-udeb is re-installed at the beginning if the installer process, often causing crashes in case of version mismatch (because the libnss* libraries were in a different package). Now libc6-udeb is unpacked while building the image, but given we still have half a dozen of libraries without the corresponding udeb, mklibs copy them on the image, as well as the dependencies, which presumably includes libresolv.so.2. > It's been a while since I last looked at/understood mklibs stuff though, > feel free to fix my suspicions/conclusions. The long term solution is to package all the libraries into udeb packages. That way we can simply get rid of the mklibs pass. The workaround are to make sure the chroots are up-to-date (which should be the case now on the build daemons). An other alternative would be to avoid copying a library in mklibs if it is already present in the image. That might break if some very strict dependencies are used, though I guess the way the udebs are downloaded, they should always have the same or a newer version than in the chroot. Aurelien -- Aurelien Jarno GPG: 4096R/1DDD8C9B aurelien@aurel32.net http://www.aurel32.net
Attachment:
signature.asc
Description: PGP signature