Testsuite races

[Jeff Law <law at redhat dot com>]

As HJ mentioned back in Sep 2011 [1], tests in the intl subdirectory can 
intermittently fail because several try to update the same file (it's 
more than just tst-codeset/tst-gettext3) which is then read by the 
actual code we're testing.  If the update happens at just the right time 
the test obviously fails.

Needless to say this is quite annoying, particularly for folks that are 
new to glibc and thus not familiar with the intermittent failures and 
for anything which does automated comparisons of test runs.

It seems to me there's a few ways to deal with this.

  1. Disable -j.  Obviously this is undesirable, even when just
  applied to a particular subdir.

  2. Do some file locking in the setup scripts.  flock is rather
  gross, but it is effective and its use is fairly widespread.

  3. Introduce dependencies in the *.out targets so that they
  execute serially if there's a file dependency in their setup scripts

  4. Do the setup once for all the tests in a new make target.  Have
  each .out depend on that new make target.  Not sure if this is
  feasible or not.

I've done #2 & #3 here for testing purposes with success.  I really 
don't care what solution is selected, just that we pick one and 
eliminate the intermittent failures that waste people's time.

Thoughts?

Jeff

[1] http://cygwin.com/ml/libc-alpha/2011-09/msg00088.html