( cd libreswan-web/slave/ && make kvm-install-base-domain )# Example

See: http://testing.libreswan.org/

Manual Setup

To publish the results from running make kvm-results as a web page, point the make variable LSW_WEBDIR at the web page's HTML directory. For instance, either by adding it to Makefile.inc.local:

$ mkdir /var/www/html/results echo LSW_WEBDIR=/var/www/html/results >> Makefile.inc.local

or your .profile:

$ echo export LSW_WEBDIR=/var/www/html/results >> ~/.profile

Automated Testing

Setup

Lets assume everything is being set up under ~/libreswan-web/:

  $ mkdir ~/libreswan-web/

With the following layout:

  ~/libreswan-web/slave/ - repository/directory under test
  ~/libreswan-web/master/ - repostory/directory with scripts
  ~/libreswan-web/pool/ - directory containing VM disks et.al.
  ~/libreswan-web/results/ - directory containing published results

and optionally:

  /tmp/pool - tmpfs containing test vm disks

• check that the host machine is correctly configured, see:

  https://libreswan.org/wiki/Test_Suite_-_KVM#Preparing_the_host_machine

• create the html directory where the results will be published

For instance, to set up results/:

  $ cd ~/libreswan-web/
  $ mkdir results

• create the pool directory for storing permanent VM disk images

For instance, assuming building and testing is being run from the sub-directory libreswan-web/:

  $ cd ~/libreswan-web/
  $ mkdir -p pool/

• checkout a dedicated repository for running tests (aka slave/)

In addition to regular updates using "git fetch + git rebase", this repository is switched to the commit being tested using "git reset --hard".

  $ cd ~/libreswan-web/
  $ git clone https://github.com/libreswan/libreswan.git slave/

• configure the dedicated test repository as desired

increase the number of reboots allowed in parallel (since a reboot seems to tie up two cores a rule of thumb is number-cores/2):

  $ echo 'KVM_WORKERS=2' >> slave/Makefile.inc.local

increase the number of test domains (and give them unique prefixes so that they don't run with the default domain names):

  $ echo 'KVM_PREFIXES=w1. w2.' >> slave/Makefile.inc.local

enable the wip tests:

  $ echo "KVM_TEST_FLAGS=--test-status 'good|wip'" >> slave/Makefile.inc.local

move the test domains to /tmp (tmpfs):

  $ echo 'KVM_LOCALDIR=/tmp/pool' >> slave/Makefile.inc.local

• checkout a repository for the web sources and scripts (aka master/)

  $ cd ~/libreswan-web/ $ git clone https://github.com/libreswan/libreswan.git master/

• create the base domain (creating the base domain requires a TTY; blame kvm):

  $ cd ~/libreswan-web/ $ ( cd slave/ && make kvm-install-base-domain )

Running

Assuming results are to be published in the directory libreswan-web/results/ (see above), the testing script is invoked as:

Either:

$ cd libreswan-web/
$ rm -f nohup.out
$ nohup master/testing/web/tester.sh slave/ results/ &
$ tail -f nohup.out

or:

$ rm -f nohup.out
$ nohup ./libreswan-web/master/testing/web/tester.sh libreswan-web/slave libreswan-web/results/ &
$ tail -f nohup.out

Restarting and Maintenance

The following things seem to go wrong:

• over time, the test results can get worse

The symptom is an increasing number of "unresolved" test results with an error of "output-missing". It happens because the domain took too long (more than 2 minutes!) to boot.

tester.sh works around this by detecting the problem and then rebuilding domains, but sometimes even that doesn't work so things need to be cleaned up.

• the build crashes

For instance a compiler error, or something more serious such as as corrupt VM.

To mitigate this cascading, after a build failure, tester.sh will reset itself and wait for new changes before attempting a further build

• the disk fills up

Test result directory can be pruned without a restart. Once the current run finishes, runner.sh will re-build the web pages removing the deleted directories (you just need to wait).

Included in the restart instructions below are suggests for how to find directories that should be pruned.

If a restart is required, the following are the recommended steps (if you're in a hurry, reboot the machine then skip all the way to the end with "restart"):

• if necessary, crash the existing runner.sh:

while killing runner.sh et.al. works, it is easier/quicker to just crash it by running the following a few times:

  $ cd libreswan-web/
  $ ( cd slave/ && make kvm-uninstall )

• (optional, but recommended) upgrade and reboot the test machine:

  $ sudo dnf upgrade -y $ sudo reboot

• (optional) cleanup and update the slave (tester.sh will do this anyway)

  $ cd libreswan-web/ $ ( cd slave/ && git clean -f ) $ ( cd slave/ && git pull --ff-only )

• (optional) update the master repository:

Remember to first check for local changes:

  $ cd libreswan-web/
  $ ( cd master/ && git status )
  $ ( cd master/ && git pull --ff-only )

• (optional) examine, and perhaps delete, any test runs where tests have 'missing-output':

  $ cd libreswan-web/ $ grep '"output-missing"' results/-g-*/results.json | cut -d/ -f1-2 | sort -u

• (optional) examine (and perhaps delete) test runs with no results.json:

  $ cd libreswan-web/ $ ls -d results/-g-*/ | while read d ; do test -r $d/results.json || echo $d ; done

• (optional) examine, and perhaps delete, some test results:

• use gime-work.sh to create a file containing, among other things, a list of test runs along with their commit and "interest" level (see below):

  $ ./master/testing/web/gime-work.sh results slave 2>&1 | tee commits.tmp
  

• strip the raw list of everything but test runs; also exclude the most recent test run (so the latest result isn't deleted):

  $ grep TESTED: commits.tmp | tail -n +2 | tee tested.tmp
  

• examine, and perhaps delete, the un-interesting (false) test runs

  Un-interesting commits do not modify the C code and are not a merge point. These are created when HEAD, which is tested unconditionally, isn't that interesting. Exclude directories already deleted.

  $ awk '/ false$/ { print $2 }' tested.tmp | while read d ; do test -d "$d" && echo $d ; done
  

• examine, and perhaps delete, a selection of more interesting (true) test runs.

  More interesting commits do modify the C code but are not a merge. Exclude directories already deleted.

  $ awk '/ true$/ { print $2 }' tested.tmp | while read d ; do test -d "$d" && echo $d ; done | shuf | tail -n +100
  

• start <tt>tester.sh</tt>, see above

Rebuilding

From time-to-time the web site may require a partial or full rebuild.

For HTML (.html, .css and .js) files, the process is straight forward. However, for the .json files, the process can be slow (and in the case of the results, a dedicated git repository is needed).

• create a repository for rebuilding the web site (aka scratch/)

When re-generating the results from a test run (for instance as part of rebuilding the web-site after a json file format change), this repository is "git reset --hard" to the original commit used to generate those results.

For instance, to set up libreswan-web/scratch/:

  $ cd libreswan-web/
  $ git clone https://github.com/libreswan/libreswan.git scratch/

• make web [WEB_SCRATCH_REPODIR=.../libreswan-web/scratch]

Update the web site.

If WEB_SCRATCH_REPODIR is specified, then the result.json files in the test run sub-directories under $(WEB_SUMMARYDIR) are also updated.

• make web-results-html

Update any out-of-date HTML (.html, .css and .json) files in the results sub-directories.

Touching the source file testing/web/results.html will force an update.

• make web-commits-json

Update the commits.json file which contains a table of all the commits. Slow.

Touching the script testing/web/json-commit.sh, which is used to create the files, will force an update. So too will deleting the .../commits/ directory.

• make web-results-json WEB_SCRATCH_REPODIR=.../libreswan-web/scratch

Update the results.json file in each test run's sub-directory. Very slow. Requires a dedicated git repository.

Touching the script testing/utils/kvmresults.py, which is used to generate results.json, will force an update.

• make '$(WEB_SUMMARYDIR)/<run>/results.json' WEB_SCRATCH_REPODIR=.../libreswan-web/scratch

Update an individual test run's results.json file. Slow. Requires a dedicated git repository.

Nuances

Some things, annoyingly, don't quite work right:

• comparisons sometimes loose a result

The code fetching all the json result tables is likely racy, when too may results are clicked too fast a result goes missing. The work around is to de-select and then re-select the missing result.

• libreswan's repo contains "knots"

When following a linear history (parent links), the commit dates normally decrease. In this repo they some times increase resulting in graph doubling back on itself.

• libreswan's repo contains "plats"

As a generalization, is good to merge local changes onto the remote first-parent branch and not the reverse. Unfortunately git pull (git merge) does the reverse by default. The result is that first-parent keeps switching with development branches.

• clicking links toggls a results selection

For instance, control-click a result's subject hash link (to open a new tab displaying the diff)) also toggls the results appearance under the comparison tab

Improvements

Some things could work better:

• "good" and "wip" each list their errors separately

So it is easier to identify "good" errors.

Unfortunately, the raw data (found in results.json) isn't currently included in summary.json.

• examine (compare) individual tests

For instance, select a test in the "Compare Results" tab would display (graph?) that test's history under a "Compare Test" tab.

The raw test results are in <run>/<test>/OUTPUT/result.json.

• Bookmark comparisons

For instance results?hash,hash.

The required information seems to be in location.search.

• Use an accumulative bar graph instead of a scatter plot

It probably better represents the data. However, since it needs to also plot branches things could get interesting.

• Graph should pan and zoom

• trim older directories so the total is kept reasonable

• use rowspan=... in the summary table