Infinite loop when using mutually-referential constants

[jeremyevans]

The following code appears to cause an infinite loop in JRuby 1.7.17 (OpenBSD) and 1.7.6 (Windows):

def a
  c = Class.new
  c2 = Class.new
  c.const_set(:A, c2)
  c2.const_set(:A, c)
end

Calling that a method causes the CPU to go to 100%, and requires a kill -9 to kill the process.

[headius]

Weird.

[headius]

This one is a little tricky...there's some differences in constant-definition logic in JRuby. The problem is basically that when doing const_set, we do an additional step that sets the parent of a module/class that doesn't currently have a parent. MRI does not appear to have this logic, but if I remove it from setConstantCommon, a number of specs that do want to see parent updated properly no longer pass.

[headius]

Here's my diff. This patch isn't right and this fix isn't going to make 1.7.18 because it's too risky to change this day before release.

diff --git a/core/src/main/java/org/jruby/RubyModule.java b/core/src/main/java/org/jruby/RubyModule.java
index c17f30e..19b2d20 100644
--- a/core/src/main/java/org/jruby/RubyModule.java
+++ b/core/src/main/java/org/jruby/RubyModule.java
@@ -3354,7 +3354,6 @@ public class RubyModule extends RubyObject {
             RubyModule module = (RubyModule)value;
             if (module != this && module.getBaseName() == null) {
                 module.setBaseName(name);
-                module.setParent(this);
             }
         }
         return value;

[headius]

For anyone that attempts this bug... more facts.

• Anonymous classes in Ruby normally do get the first name they're assigned to, but it appears that only happens for a direct constant assignment. In other words, const_set does not always set the name?
• Parent assignment seems to be implicit in some operations, but not others. We need to sort out how and when MRI actually updates anonymous class/module name and parent.

[headius]

I'm looking at this casually for 1.7.20 and wondering how synthetic this is. @jeremyevans Did you run into this with real code?

[headius]

FWIW, you have to ^Z and kill because IRB traps SIGINT. If you run this at command line it can be interrupted with ^C like a normal JVM program.