"No messaging setup" in START_TEST

[krichter722]

Continuing http://sourceforge.net/p/check/bugs/111/: Running make check in an autotools-based project causes all tests to fail and check_msg.c:80: No messaging setup to be logged to test-suite.log. I found http://sourceforge.net/p/check/bugs/74/ which doesn't apply because I don't use Windows and http://sourceforge.net/p/check/mailman/message/2750522/ which doesn't contain any replies.

Minimal working example is provided at https://github.com/krichter722/check-no_messaging_setup-issue based on check's check_money example. I'm also providing a travis build result at https://travis-ci.org/krichter722/check-no_messaging_setup-issue/builds/107650108 which reproduces the issue!

I didn't manage to upload the project as a zip because travis didn't let me. I think it would be good to archive it here and for me to keep by repository list clean in the future.

[brarcher]

Thanks for coming up with a minimal example to work from!

I've forked the example and got it building for both Linux and OSX in this branch. What I've found is that the issue does not occur using Check 0.10.0. That is, if Check 0.10.0 was installed before tests were run either using apt (for Linux) or brew (for OSX) the test suite in question will compile and fail in the expected way:

Running suite(s): auto-apt
0%: Checks: 1, Failures: 1, Errors: 0
check_auto_apt.c:8:F:test_execl:test_execl:0: Assertion 'ret_value==0' failed: ret_value==1, 0==0

I attempted to use a locally built version of Check to reproduce the issue. At first I used the latest version of Check. My first attempt did find the "no messaging setup" issue, but only on Linux; OSX does not hit the issue. However, I realized that Check was still being installed prior to building it locally. To remove the possibility that the wrong version was being used, I then attempted to not install Check and only build locally. Using both revision 0.10.0 and the latest master branch, the issue did not occur.

Note though that when building locally I needed to make the following adjustments to the Makefile for the unit tests:

 sed 's/check_pic/check/g' tests/Makefile.am | sed 's/-lrt//g' > tests/Makefile.am.tmp
 mv tests/Makefile.am.tmp tests/Makefile.am

Namely, I used the dynamically built library which was locally built instead of the static library which is available on Ubuntu. This build shows building the latest Check locally, linking using -lcheck, and the test failing in the expected way. This was the .travis.yml configuration script which was used.

I do not know the root issue, as I've not seen the issue locally and cannot poke around in detail. However, I wonder about two things:

1. The configure script used in the minimal example is using a macro to find Check, and flags are being added manually in the Makefile. The latest version of Check provides an example which uses pkg-config to find Check and the necessary flags. Switching to that configure.ac file and Makefile style may help the example build with the proper flags, if any are missing or some are unnecessary.
2. Perhaps there is something strange with the static library provided by Ubuntu, and maybe not all of the necessary flags are added to the hand generated Makefile in order to get everything to work correctly.

If you are up for it, could you attempt to update the configure.ac and Makefile.am files to leverage pkg-config more and see if the issue still happens using either a version of Check built locally or the Ubuntu provided static library?

[krichter722]

It's uncommon (in my experience) that the installation from source requires manual adjustment (see https://travis-ci.org/krichter722/check-no_messaging_setup-issue/builds/119692062 without Ubuntu-provided check). Is that an issue (regardless of usage of pkg-config)? Should it be fixed before investigating further?

The

sed 's/check_pic/check/g' tests/Makefile.am | sed 's/-lrt//g' > tests/Makefile.am.tmp
mv tests/Makefile.am.tmp tests/Makefile.am

fix works with 0.10.0-69-g82cdb29 (see https://travis-ci.org/krichter722/check-no_messaging_setup-issue/builds/119725269) (had to add #include <stdint.h> to the test case). Guess this can be considered a user error, but the feedback is still strange and could be improved. What is -lcheck_pic anyway? Thanks for your support.

[brarcher]

What is -lcheck_pic anyway?

I'm not really sure. It appears this is what Ubuntu choose to name the Check library when it was build statically.

Is that an issue (regardless of usage of pkg-config)? Should it be fixed before investigating further?

It seems to be that the issue is how the linker dependencies are generated. When the static library is used (check_pic) additional dependencies are necessary which were missing from the Makefile. pkg-config should report the necessary flags.

Why the "no messaging" issue can occur is still a mystery to me.

[krichter722]

I still have a lot of trouble with the reproduction of no messaging setup. I can trigger it now outside START_TEST with https://github.com/krichter722/check-no_messaging_setup-issue/commit/2744b0897ce00abb022cc5209ca8077796ccecdf (predecessor works). The symptoms were the same within START_TEST before. I guess one can expect that it doesn't work outside START_TEST, but maybe the feedback could be improved, so that one might say the issue isn't reproducible. Besides improving feedback it'd be good to extract one or two test cases as well.

I'm not very good with autotools and couldn't figure out a bunch of compilation errors which made me add a lot of extranous files to the (no longer) minimal example.

[andr-dots]

I've had such an issue, too. It comes from build (execute) folder permissions. I didn't had rights to create files while running tests. And I've figured out what is wrong not so fast...

I've quickly checked the code and open_tmp_file is just returns NULL if it could not create the file. I think this is the case. Moreover, it tries to create file in current directory (.) in my system. If it is a temporary file I think it should be created in the /tmp folder.

[brarcher]

I think that the open_tmp_file() function attempts to open a temporary file in one of several different places. If the system does not have mkstemp() defined (e.g. Windows) then tempnam() is used. This results in the following heuristic for determining where to put the temporary file:

1. If the TMPDIR environment variable is defined, use that directory
2. If the TEMP environment variable is defined, use that directory
3. Use the platform defined temporary directory

For system with mkstemp() defined (e.g. Linux) the following heuristic is used:

1. If the TEMP environment variable is defined, use that directory
2. Use the current directory

This could be updated to better map with what Windows would do. Namely, check if TMPDIR is defined first, then fallback to TEMP, then use the current directory.

[brarcher]

@krichter722, could you see if the change in the following merge request resolves the issue you are seeing?

#81

[andr-dots]

I've checked on my system, TMPDIR is not set. But P_tmpdir macro returns /tmp folder. And in case if this macro doesn't exist on linux systems, /tmp could be used because it is a temporary directory from FHS.

[brarcher]

I'll add a check for P_tmpdir as well. That should cover all the bases.

[brarcher]

A fix for the permission issue has been merged. Kindly re-test to determine if it resolves the issue.

[brarcher]

Due to no response after a fix has been submitted, I'll close this issue. If the issue is not resolved in the master branch, let me know.

[Earnestly]

Hi, I appear to be getting a similar error when attempting to build the
libinput project.

When asked to view the tests/test-suite.log file I find the "No messaging setup" message which appears to be causing the tests to fail:

=========================================
   libinput 1.7.0: test/test-suite.log
=========================================

# TOTAL: 3
# PASS:  2
# SKIP:  0
# XFAIL: 0
# FAIL:  1
# XPASS: 0
# ERROR: 0

.. contents:: :depth: 2

FAIL: libinput-test-suite-runner
================================

check_msg.c:80: No messaging setup
FAIL libinput-test-suite-runner (exit status: 2)