Issue #5063: Module defaults need to be considered and copied automatically during recursive dependency solving
Modules are incorrectly copied when their artifacts shadow the names of non-modular RPM dependencies
When Pulp tries to copy a module like openmpi or mariadb from the Fedora modular repository (https://dl.fedoraproject.org/pub/fedora/linux/updates/30/Modular/x86_64/), using either the recursive or recursive-conservative depsolving strategies, it will additionally (and incorrectly) copy modules like perl, perl-bootstrap, and setools.
This happens because one of the artifacts in these modules (modules: openmpi and mariadb) depends on the perl package, which is meant to be provided by the normal Fedora repositories. However in this case there also exists a "perl" module, and libsolv tries to use artifacts from the "perl" module to satisfy the dependency. This should not happen because the openmpi and mariadb modules do not not declare a modular dependency in the "dependencies" section of their modulemd documents.
Dependency solving uses a libsolv "pool" that contains the entire contents of the source repository, and in conservative mode, also the entire contents of the target repository. We need to change the behavior of the depsolver such that such that all non-modular RPMs in the pool are "considered" during depsolving, but only the modular RPMs that are part of the module being copied or an explicitly listed modular dependency are "considered", and all others are excluded from depsolving.
I'm unsure if a mechanism to do hide certain units from the depsolver is exposed to us, see: https://github.com/openSUSE/libsolv/issues/258.
If not, it may be necessary to hack around the problem by being more selective about what module / modular artifact solvables we load into the pool to begin with.
--- document: modulemd version: 2 data: name: openmpi stream: 4.0 version: 3020190331170644 context: a5b0195c arch: x86_64 summary: Open Message Passing Interface description: >- Open MPI is an open source, freely available implementation of both the MPI-1 and MPI-2 standards, combining technologies and resources from several other projects (FT-MPI, LA-MPI, LAM/MPI, and PACX-MPI) in order to build the best MPI library available. A completely new MPI-2 compliant implementation, Open MPI offers advantages for system and software vendors, application developers, and computer science researchers. For more information, see http://www.open-mpi.org/ . license: module: - MIT content: - BSD and MIT and Romio dependencies: - buildrequires: platform: [f30] requires: platform: [f30] references: community: https://www.open-mpi.org/ documentation: https://www.open-mpi.org/doc/v4.0/ tracker: https://github.com/open-mpi/ompi/issues profiles: buildroot: rpms: - openmpi-devel default: rpms: - openmpi api: rpms: - openmpi components: rpms: openmpi: rationale: The core package ref: 4.0 buildorder: 10 arches: [aarch64, armv7hl, i686, ppc64le, s390x, x86_64] artifacts: rpms: - openmpi-0:4.0.1-1.module_f30+3806+d6645847.src - openmpi-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - openmpi-debuginfo-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - openmpi-debugsource-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - openmpi-devel-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - openmpi-devel-debuginfo-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - openmpi-java-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - openmpi-java-devel-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - python2-openmpi-0:4.0.1-1.module_f30+3806+d6645847.x86_64 - python3-openmpi-0:4.0.1-1.module_f30+3806+d6645847.x86_64 ...
See: no "dependencies" listed beyond the standard "platform"
The openmpi-devel package is the one that has a requires on /usr/bin/perl
Below is a conversation I had with Igor Gnatenko where he mentioned that the approach described above is flawed and suggested an alternative approach. That is what I will be implementing.
<ignatenkobrain> dalley: yes, you need to know about defaults <ignatenkobrain> so when you are resolving deps, you need to take into account: <ignatenkobrain> * non-modular content <ignatenkobrain> * default module streams (even if they have lower version of non-modular content, they should be prefered) <ignatenkobrain> * all combinations of you modular dependencies <ignatenkobrain> dalley: so I think you do it wrong when you enable / disable some content. You should enable everything by default and then set up libsolv dependencies in a proper way <ignatenkobrain> and then handle some special cases <dalley> ignatenkobrain, how do we do that? we have to load "perl-4:5.28.2-436.module_f30+4677+de2e7282.x86_64" (or whatever) into the depsolver, how do we tell it "this is a modular RPM, don't try to use it to satisfy the dependency of "mariadb-3:10.4.2-1.beta.module_f30+2924+d4dbc42b.x86_64" <ignatenkobrain> dalley: from which module is that perl coming? <ignatenkobrain> is it default stream of perl module? <ignatenkobrain> if not, does mariadb module depend on it? <dalley> mariadb doesn't depend on the perl module at all, but it does depend on perl, and "perl-4:5.28.2-436.module_f30+4677+de2e7282.x86_64" satisfies that so the depsolver tries to use it to satisfy the dep <dalley> if it didn't exist at all, it would come back and say "we have a problem, perl is missing" and so we would fake it <dalley> because DNF may know about all these extra repos, but Pulp doesn't <ignatenkobrain> dalley: well, that's the thing... perl's module stream is not default <ignatenkobrain> so it is in conflict <ignatenkobrain> let me find some relevant code in fus <ignatenkobrain> dalley: https://github.com/fedora-modularity/fus/blob/master/fus.c#L1290-L1318 <ignatenkobrain> dalley: so for default modules you need to add Provides: module-default() and then you can easily query default and non-default modules. <ignatenkobrain> then you need to disable irrelevant non-default modules <ignatenkobrain> which would make some constraints on packages from those modules to not be installable <ignatenkobrain> and that will pull non-modular perl package :0 <ignatenkobrain> because that modular perl requires module(perl:5.26) or whatever, so if all providers of that capability are disabled, it won't be chosen to satisfy dependency
PR where future fixes will land is here: https://github.com/pulp/pulp_rpm/pull/1383
Please register to edit this issue