From: David Masover <ninja@...>
Date: 2009-07-23T13:44:30+09:00
Subject: Garbage collection oddities

Given this code:

class WatchMeDie
  def self.finalizer
    proc do |id|
      puts "Finalized #{id}"
    end
  end
  def initialize
    ObjectSpace.define_finalizer self, &WatchMeDie.finalizer
  end
end

a = WatchMeDie.new
a = nil
GC.start
puts 'end of program'

# # # 

This produces the expected outcome on 1.8.7:

Finalized 69858803067800
end of program

On 1.9.1, however, it doesn't:

end of program
Finalized 8054060

That is, on 1.9.1, the object is never collected. However, if I create two 
such objects:

a = WatchMeDie.new
b = WatchMeDie.new
a = nil
b = nil
GC.start

Once I do this, the first object is collectible. It seems the most recent 
WatchMeDie object will never be collected before the program ends.

Not that any of this really matters, but it complicates what I was actually 
testing:

def make_a_thread
  Thread.new do
    sleep 1000
  end
end

def test_run
  make_a_thread
  WatchMeDie.new
end

a = test_run
b = WatchMeDie.new
a = nil
GC.start
puts 'end of program'


This example works as expected -- one WatchMeDie is finalized before program 
end, and one after. However, if we change test_run to:

def test_run
  make_a_thread
  foo = WatchMeDie.new
end

1.9.1 still works as expected. However, 1.8 does not -- because I've now 
assigned that object to a local variable, even one that didn't exist when I 
called make_a_thread, it now can't be garbage collected.

One possible workaround, doesn't:

MyThreadProc = proc do
  sleep 1000
end
def make_a_thread
  Thread.new &MyThreadProc
end

In other words, for some bizarre reason, either the thread or the proc still 
cares enough about that local scope, and the scope of all its callers, to hold 
onto them for posterity.

I've worked around this in 1.8 by sending the thread spawning itself to 
another thread, one created with a somewhat cleaner scope and call stack. But 
it's a brutally ugly hack, and there have to be performance implications.

Fortunately, 1.9.1 appears to do the right thing, here. Unfortunately, this 
kind of stuff is difficult to test. So I'm curious: Is this unique to 1.8.7, or 
does it exist in 1.8.6, also? And is it likely to be fixed, or should I just 
strongly encourage people to upgrade to 1.9?