#1042978 Tests invalid package upgrades

Package:
release.debian.org
Source:
release.debian.org
Submitter:
Ben Hutchings
Date:
2023-11-30 13:30:02 UTC
Severity:
normal
#1042978#5
Date:
2023-08-03 16:01:36 UTC
From:
To:
ci.debian.net is running dkms's autopkgtest with packages from testing
except for linux's binaries updated from unstable.  The test fails
because linux-headers-amd64 is uninstallable in this scenario:

* The available version of linux-headers-amd64 will be the latest,
  6.4.4-2.  It depends on (currently) linux-headers-6.4.0-1-amd64
  at the same version.
* The available version of linux-headers-6.4.0-1-amd64 (built from
  src:linux-signed-amd64) will be 6.4.4-1.

This cross-source-package version lock is unusual but has been
implemented deliberately to ensure that linux and linux-signed-*
always migrate from unstable to testing together.

ci.debian.net must therefore also test updating the binaries from all
these source packages together, not just those built from linux.  If
it's not possible to do this in a general way then we need a specific
workaround that can be used in dkms and any other affected packages
to prevent this invalid test scenario.

Ben.
#1042978#10
Date:
2023-08-03 16:06:17 UTC
From:
To:
The test log for reference:
https://ci.debian.net/data/autopkgtest/testing/amd64/d/dkms/36392756/log.gz

Ben.
#1042978#19
Date:
2023-08-03 20:01:39 UTC
From:
To:
Hi Ben,

But ci.debian.net just does what it's been asked to do by the client, in
this case britney. So, if anything it's britney2 that needs to be
changed to support this. But, britney2 tries to deduce what needs to be
tested together from the package relations. Normally, what you're seeing
here is the result of a test where the signed packages haven't been
build yet. britney retries tests after 24 hours and normally it retries
with linux-signed-* in the list of pinned packages as you can see in the
dkms history. The question that now arises is why it doesn't do that now.

Paul
#1042978#26
Date:
2023-08-05 00:15:59 UTC
From:
To:
Thank you for redirecting this appropriately.

I suspect that the difference is that most linux updates bump ABI (thus
changing package names) and this last update did not need to.  The
package that linux-headers-amd64/unstable depends on typically don't
exist in testing at all, but in this case it does (at the wrong
version).  Perhaps something in britney's invocation of debci isn't
paying attention to all the version constraints.

Ben.