@GrahamcOfBorg build rcs

I've pushed a the corrected hardening configuration, but haven't yet managed to reproduce the t999 failure seen in ofborg.

@GrahamcOfBorg build rcs

I'll see if I can get the test case to print what failed, and then we can see what is failing under ofborg.

This latest version includes verbose reporting on test failure. Let's see what's happening.

Interesting: for commit f11fd5d4e8801df18e8a5a8a823332b16b700a3c, on my macOS 10.12 machine, I observed the following:

• nix-build -A rcs . fails [1]: https://gist.github.com/strager/a6112c53167baa24a49abe51374cedd8
• nix-shell -A rcs --pure then genericBuild successfully builds and run tests [2]: https://gist.github.com/strager/ec14fbc7655fbf4f447f9850243e41ee

[1] with both sandbox=true and sandbox=false
[2] but then fails when installing, which is normal

I have found reason the tests fail: there's an observable difference in the build environment.

The difference can be seen with

nix-build --expr 'with (import ./. {}); runCommand "test-getlogin" {} 
    "${ruby}/bin/ruby -r etc -e \"puts Etc.getlogin\" "'

Where rcs tests pass, this returns nixbld1, and where rcs tests fail, this returns root. I don't know the cause of the difference.

The test failure is a result of two different methods of determining the username. In rcs, it looks at, in order, $LOGNAME, $USER, getlogin, getpwnam(ruid()). The test scripts look at, in order, id -un, whoami, $LOGNAME and $USER. In the build environments I inspected, LOGNAME and USER are unset. Ultimately rcs binaries use getlogin, and the test driver uses id -un. When it comes to testing expectations, the file is expected to be locked by the current user "nixbld1", but instead is locked by "root". We can make rcs build by setting USER to the result of id -un.

This exact situation appears to have been observed in 2014: https://lists.gnu.org/archive/html/hydra-users/2014-12/msg00002.html

We can make rcs build by setting USER to the result of id -un.

Possible upstreamable solution: Remove the call to getlogin_r in getusername (src/b-excwho.c). This could fix other cases too, like 'sudo -u anotheruser rcs ...'. (Or maybe not, and it'll break other things. I didn't look or think too carefully. =X)

I'm glad you were able to find the cause for the build weirdness. I guess nix-shell --pure isn't so pure after all. ;P