When I rewrote the compute_scc function into an iterative variant, I
almost included a bug that could make it come up with components that
were not strongly-connected.
Signed-off-by: Niels Thykier <niels@thykier.net>
Rewrite _compute_scc to be iterative to avoid call recursion limit for
graphs with long dependency chains.
Signed-off-by: Niels Thykier <niels@thykier.net>
Sometimes the solver would correctly compute which items are grouped
together, but then fail to emit some groups. This is trivially
fixable by forcing an update of "before" + "after" relations.
Admittedly, this looks unnecessary, so it may just be hiding the bug.
Nonetheless, the change makes the new test_solver_simple_scc test
produce the expected result.
Signed-off-by: Niels Thykier <niels@thykier.net>
Britney has a special case for essential packages to ensure that any
package that with essential packages are not installable. This check
did not account for a case, where a package is not co-installable with
two or more pseudo-essential package part of the same OR dependency.
A contrived example based on real world data:
Package: foo
# Conflict with all providers of "awk"
Conflicts: mawk | gawk | original-awk
This alone is actually not sufficient to trigger the bug, as
_get_min_pseudo_ess_set is in theory some times smart enough to pick
an "obvious" solution between the pseudo-essential option.
When it does, one of the above ends up in the (de-facto) essential set
and then the installability tester correctly rejects "foo".
Though, even with the fix above, the handling for this is probably not
correct if the essential set is not (fully co-)installable. However,
that basically only happens if we are bootstrapping an architecture
(or testing is royally broken, in which case this is the least of our
worries).
Signed-off-by: Niels Thykier <niels@thykier.net>
Niels Thykier