I think this was fixed by Nikolay's latest (rev d27bbaf93422).

But please test it for yourself.

yes it is, but I think the fix introduces a race condition. The while loop 
should not be removed (POSIX signals are not queued).

hmm, i dont see a race condition. Could you explain more?

If two children terminate in the same time, SIGCHLD may be raised only once.

This condition can be triggered easily when running some code through the C API 
because the signal handler is called only after the C function returns.

The attached script demonstrates it.

With 44e4c8257a9d it outputs:
  first finished 3.0
  second finished 3.0
  end 4.0

With d27bbaf93422 it outputs:
  first finished 3.0
  end 4.0
(the termination of the second process not reported)

OK, reopening.

Could it be that this is also causing the buildbot failures that are being 
reported?  E.g. http://bugs.python.org/issue19294

to me it looks unrelated because test_subprocess_shell starts only one 
subprocess

I believe this was fixed in CPython by 
http://hg.python.org/cpython/rev/28397ecf2316 . Please verify that your latest 
test works with CPython 3.4alpha4 (replace tulip with asyncio in the test) -- 
it has a dependency that I haven't got.

I will port that fix to the Tulip repo once things settle a bit.

I confirm. Both tests work well now

Thanks! Closing again. (The Tulip port has happened.)