Run commands with unique pgid to improve timeout cleanup

To ensure that all child processes are properly signaled to exit before
forcibly killing them, use `setsid()` in the child process to give the
child (and any grandchildren it might create) a new and unique process
group. If the command times out, kill and reap the entire process group.

1 files changed, 52 insertions, 1 deletions

diff --git a/spec/mixlib/shellout_spec.rb b/spec/mixlib/shellout_spec.rb
index 5e92a54..582aef4 100644
--- a/spec/mixlib/shellout_spec.rb
+++ b/spec/mixlib/shellout_spec.rb
@@ -825,7 +825,7 @@ describe Mixlib::ShellOut do
         end
       end
 
-      context 'with subprocess that takes longer than timeout', :focus do
+      context 'with subprocess that takes longer than timeout' do
         def ruby_wo_shell(code)
           parts = %w[ruby]
           parts << "--disable-gems" if ruby_19?
@@ -889,6 +889,57 @@ describe Mixlib::ShellOut do
 
           end
         end
+
+        context "and the child process forks grandchildren" do
+          let(:cmd) do
+            ruby_wo_shell(<<-CODE)
+              STDOUT.sync = true
+              trap(:TERM) { print "got term in child\n"; exit!(123) }
+              fork do
+                trap(:TERM) { print "got term in grandchild\n"; exit!(142) }
+                sleep 10
+              end
+              sleep 10
+            CODE
+          end
+
+          it "should TERM the wayward child and grandchild" do
+            # note: let blocks don't correctly memoize if an exception is raised,
+            # so can't use executed_cmd
+            lambda { shell_cmd.run_command}.should raise_error(Mixlib::ShellOut::CommandTimeout)
+            shell_cmd.stdout.should include("got term in child")
+            shell_cmd.stdout.should include("got term in grandchild")
+          end
+
+        end
+        context "and the child process forks grandchildren that don't respond to TERM" do
+          let(:cmd) do
+            ruby_wo_shell(<<-CODE)
+              STDOUT.sync = true
+              trap(:TERM) { print "got term in child\n" }
+              fork do
+                trap(:TERM) { print "got term in grandchild\n" }
+                sleep 10
+              end
+              sleep 10
+            CODE
+          end
+
+          it "should TERM the wayward child and grandchild, then KILL whoever is left" do
+            # note: let blocks don't correctly memoize if an exception is raised,
+            # so can't use executed_cmd
+            lambda { shell_cmd.run_command}.should raise_error(Mixlib::ShellOut::CommandTimeout)
+            shell_cmd.stdout.should include("got term in child")
+            shell_cmd.stdout.should include("got term in grandchild")
+
+            # A little janky. We get the process group id out of the command
+            # object, then try to kill a process in it to make sure none
+            # exists. Trusting the system under test like this isn't great but
+            # it's difficult to test otherwise.
+            lambda { Process.kill(:INT, shell_cmd.send(:child_pgid)) }.should raise_error(Errno::ESRCH)
+          end
+
+        end
       end
 
       context 'with subprocess that exceeds buffersize' do