[Gluster-infra] About the recent regression test breakages

[Justin Clift]

(re-sending after subscribing to the list)

On 22/02/2014, at 6:42 PM, Vijay Bellur wrote:
> On 02/22/2014 02:54 PM, Niels de Vos wrote:
>> Hi all,
>> 
>> we recently have had an issue where regression tests were failing with
>> strange errors like 'C compiler cannot create execytables'. Also, the
>> nightly builds return the same error... Of course, when running the
>> tests locally, this works all fine.
>> 
>> The command that ./configure creates looks like this (from config.log):
>> 
>>    gcc   /d/build/install/lib/ conftest.c  >&5
>> 
>> This definitely is an incorrect command.
>> 
>> Now, because both the regression test and nightly builds fail with the
>> same error, the problem seems to be system-wide on build.gluster.org.
>> The nightly builds run under my user-account, and only create a tarball
>> and src.rpm which gets build on a different system. Running ./configure
>> as my user, also has /d/build/install/lib/ in the gcc commandline. This
>> is surely unexpected.
>> 
>> It seems that there are two global environment variables set with this
>> path. LD_LIBRARY_PATH and LDFLAGS.
>> 
>>    LD_LIBRARY_PATH:
>>    Setting a system-wide LD_LIBRARY_PATH is quite fragile. This
>>    environment variable affects the loading of dynamically linked
>>    libraries. The normal operating system services should NEVER try to
>>    load any library from /d/build/install/lib which gets build during
>>    each regression test. If the reviewer who starts the regression test
>>    does not pay attention to the change, malicious libraries could be
>>    build and installed in /d/build/install/lib, which then could affect
>>    standard operating system behaviour.
>>    LD_LIBRARY_PATH=/d/build/install/lib/ was set in
>>    /etc/profile.d/jenkins.sh. This has now been commented out, and
>>    a note in the file has been left behind for future reference. It
>>    would be much better to only set LD_LIBRARY_PATH in the scripts that
>>    execute the regression test runs (after building).
>> 
>>    LDFLAGS:
>>    Of course, LDFLAGS immediately affects any standard 'make' commands, and
>>    also ./configure uses CFLAGS and LDFLAGS from the environment.
>>    LDFLAGS=/d/build/install/lib/ was set (not a valid LDFLAGS option
>>    string) in /etc/environment. This has now been commented out, and
>>    a note in the file has been left behind for future reference.
>> 
>> 
>> I have not restarted Jenkins (don't know how to do that). With the above
>> changes the nightly builds work again, but regression tests still fail.
>> I expect that there is a Jenkins process that has LDFLAGS still set in
>> its environment.
>> 
>> Two pending actions:
>> 
>> 1. update the regression test script(s) to set LD_LIBRARY_PATH
>> 2. restart the Jenkins service to clean the environment variables
>> 
>> It would be much appreciated if system-wide changes like these at least
>> have some notes in the configuration files on why these settings are
>> needed, and who changed this. An email to this list would help others in
>> case suddenly strange errors occur...
> 
> Thanks, Niels for identifying this one.

Fantastic stuff Niels! :)

> I am the culprit here. These modifications to the build environment were meant to be temporary but I forgot to unset this after I was done debugging :-(.
> 
> My sincere apologies for all the inconvenience and I'll ensure that I never attempt such kludges without notifying this list. A beer on me to both you & Justin when I meet you next!

Nah, Niels did the work not me.  Give him mine as well. :)


> I have attempted to get both pending actions right. Please let me know if you still encounter problems.

Thanks Vijay. :>

+ Justin

--
Open Source and Standards @ Red Hat

twitter.com/realjustinclift