Use blocking waitpid only on ENOENT

This is a compromise that prevents us from leaving zombie processes
around after a failed exec raises ENOENT, but prevents us from blocking
(possibly forever) on timed-out commands. The latter case can cause
zombies, so this is not optimal for all cases. See MIXLIB-16 for a
proposed enhancement to solve that case.

1 files changed, 10 insertions, 2 deletions

diff --git a/lib/mixlib/shellout/unix.rb b/lib/mixlib/shellout/unix.rb
index af3e421..d7b8b92 100644
--- a/lib/mixlib/shellout/unix.rb
+++ b/lib/mixlib/shellout/unix.rb
@@ -82,10 +82,18 @@ module Mixlib
           end
         end
         self
-      rescue Exception
-        # do our best to kill zombies
+      rescue Errno::ENOENT
+        # When ENOENT happens, we can be reasonably sure that the child process
+        # is going to exit quickly, so we use the blocking variant of waitpid2
         Process.waitpid2(@child_pid) rescue nil
         raise
+      rescue Exception
+        # For exceptions other than ENOENT, such as timeout, we can't be sure
+        # how long the child process will live, so we use the non-blocking
+        # variant of waitpid2. This can result in zombie processes when the
+        # child later dies. See MIXLIB-16 for proposed enhancement.
+        Process.waitpid2(@child_pid, Process::WNOHANG) rescue nil
+        raise
       ensure
         # no matter what happens, turn the GC back on, and hope whatever busted
         # version of ruby we're on doesn't allocate some objects during the next