Unity continuous-integration fails in armhf due possibly to a libegl issue

[Christopher Townsend]

Since mesa 10.0 has been released, our Unity continuous-integration tests are failing on armhf builds. This is a snippet of the log:

libEGL warning: DRI2: failed to authenticate
Segmentation fault (core dumped)

The log can be found here for reference: https://jenkins.qa.ubuntu.com/view/Trusty/view/All/job/unity-trusty-armhf-ci/56/console

I don't have the core file or have access to an armhf machine to reproduce, so I'm only going on evidence based on the warning and the timing of the issue.

*Note* This is blocking automerging of Unity code, so this issue is pretty critical.

[Maarten Lankhorst (mlankhorst)]

Do you have a log of the last passing test then?

[Christopher Townsend (townsend)]

Sure do: https://jenkins.qa.ubuntu.com/job/unity-trusty-armhf-autolanding/40/consoleFull

This was on Dec. 9 which had Mesa version 9.2.2-1ubuntu2.

[Maarten Lankhorst (mlankhorst)]

Yeah but that can fail for any number of reasons, even things in unity.

Can you confirm this on x86 with LIBGL_ALWAYS_SOFTWARE=1 set in the environment? That should force a fallback on llvmpipe.

[Christopher Townsend (townsend)]

I'm quite certain that this is not a Unity issue as completely separate merge proposals all fail armhf CI. However, we don't have definitive proof of that, so we are trying other ways to debug this.

We tried the LIBGL_ALWAYS_SOFTWARE=1 env var in an amd64 build, but no where in the log does it say it's using llvmpipe for sure. The build and tests passed as they did before, so this is not a good test.

Timo A. also suggested trying a test Mesa package reverting the arm llvmpipe support. We are in the middle of getting a PPA setup to get an armhf build and then try that. We'll report back once this is done.

[Stephen M. Webb (bregma)]

Unity built and passed all tests on armhf with the reconfigured Mesa from the PPA.

http://s-jenkins.ubuntu-ci:8080/job/unity-trusty-armhf-ci/61/consoleFull

[Christopher Townsend (townsend)]

For reference, the reconfigured Mesa is just reverting this change: http://bazaar.launchpad.net/~ubuntu-branches/ubuntu/trusty/mesa/trusty/revision/206/debian/rules

[Maarten Lankhorst (mlankhorst)]

I would prefer not to revert it for now, but instead fix the bug. I've made a trusty chroot for testing on my pandaboard, how do I run those tests myself?

[Stephen M. Webb (bregma)]

To build Unity and run the tests in an up-to-date Trusty build environment you need to use the GTK fix branch (API changes in GTK prevent Unity from building using trunk).

# apt-get build-dep unity
# bzr branch lp:~townsend/unity/fix-gtk-build-error/
# cd fix-gtk-build-error
# dpkg-buildpackage

The unit tests are run as a part of the packaging build.

The file tests/dummy-xorg-test-runner.sh is a shell script use to invoke a headless X.org server against which the tests get run. I believe this is where the problem lies. It looks like the headless X.org server does not start properly on armhf targets with the new Mesa. It may be possible to avoid building Unity to reproduce this problem if it occurs just running that script.

[Launchpad Janitor (janitor)]

This bug was fixed in the package mesa - 10.0.1-1ubuntu2

---------------
mesa (10.0.1-1ubuntu2) trusty; urgency=medium

  [ Matthias Klose ]
  * Build using llvm-3.4.
 -- Maarten Lankhorst <email address hidden> Thu, 19 Dec 2013 19:21:03 +0100