First stab at a weakref test suite

ruby-talk
1

Hi all,

I decided to take a stab at creating a minimal test suite for WeakRef (based on a post by Charles Oliver Nutter for JRuby). The only problem is that I get two unexpected failures. I'm not sure if it's Test::Unit fighting with me or what.

Consider the following example:

require 'weakref'
str = 'hello'
ref = WeakRef.new(str)

p ref 'hello'
GC.start
p ref 'hello'
str = nil
p ref # 'hello'
GC.start
p ref # error

I follow the same pattern in the tests, but it doesn't raise the expected error. Any ideas?

Thanks,

Dan

# tc_weakref.rb
require 'test/unit'
require 'weakref'

class TC_WeakRef < Test::Unit::TestCase
    def setup
       @ref = nil
       @str = 'hello'
       GC.enable
    end

    def test_weakref_constructor
       assert_respond_to(WeakRef, :new)
       assert_nothing_raised{ @ref = WeakRef.new(@str) }
       assert_kind_of(WeakRef, @ref)
    end

    # TODO: Figure out why last test fails
    def test_weakref
       assert_nothing_raised{ @ref = WeakRef.new(@str) }
       assert_equal('hello', @ref)

       assert_nothing_raised{ GC.start }
       assert_equal('hello', @ref)

       assert_nothing_raised{ @str = nil }
       assert_equal('hello', @ref)

       assert_nothing_raised{ GC.start }
       assert_raise(WeakRef::RefError){ @str = @ref * 3 }
    end

    def test_weakref_is_alive_basic
       assert_nothing_raised{ @ref = WeakRef.new(@str) }
       assert_respond_to(@ref, :weakref_alive?)
    end

    # TODO: Figure out why last test fails
    def test_weakref_is_alive
       assert_nothing_raised{ @ref = WeakRef.new(@str) }
       assert_equal(true, @ref.weakref_alive?)

       assert_nothing_raised{ GC.start }
       assert_equal(true, @ref.weakref_alive?)

       assert_nothing_raised{ @str = nil }
       assert_equal(true, @ref.weakref_alive?)

       assert_nothing_raised{ GC.start }
       assert_equal(false, @ref.weakref_alive?)
    end

    def teardown
       @str = nil
       @ref = nil
    end
end

2

Since the GC is conservative, references may exist longer than you expect because the item that should be collected is still referenced from the stack. I don't know of an easy way around this.

···

On Dec 20, 2007, at 21:45 PM, Daniel Berger wrote:

Hi all,

I decided to take a stab at creating a minimal test suite for WeakRef (based on a post by Charles Oliver Nutter for JRuby). The only problem is that I get two unexpected failures. I'm not sure if it's Test::Unit fighting with me or what.

Consider the following example:

require 'weakref'
str = 'hello'
ref = WeakRef.new(str)

p ref 'hello'
GC.start
p ref 'hello'
str = nil
p ref # 'hello'
GC.start
p ref # error

I follow the same pattern in the tests, but it doesn't raise the expected error. Any ideas?

3

Eric Hodel wrote:

···

On Dec 20, 2007, at 21:45 PM, Daniel Berger wrote:

I follow the same pattern in the tests, but it doesn't raise the expected error. Any ideas?

Since the GC is conservative, references may exist longer than you expect because the item that should be collected is still referenced from the stack. I don't know of an easy way around this.

This is probably the largest reason why a weakref suite will be a little tricky. My best recommendation would be to loop calling GC.start until the reference gets collected; perhaps with an upper count... "if we reach 1000 iterations and it's not collected, it isn't working".

- Charlie

4

If it isn't collected after 1, I doubt it will be collected after 1000. You'd have better luck manipulating the stack (calling methods recursively that eventually allocate things), but you're still not guaranteed that any object will be collected when you want with ruby 1.8 or 1.9.

···

On Dec 21, 2007, at 05:50 AM, Charles Oliver Nutter wrote:

Eric Hodel wrote:

On Dec 20, 2007, at 21:45 PM, Daniel Berger wrote:

I follow the same pattern in the tests, but it doesn't raise the expected error. Any ideas?

Since the GC is conservative, references may exist longer than you expect because the item that should be collected is still referenced from the stack. I don't know of an easy way around this.

This is probably the largest reason why a weakref suite will be a little tricky. My best recommendation would be to loop calling GC.start until the reference gets collected; perhaps with an upper count... "if we reach 1000 iterations and it's not collected, it isn't working".

5

I took a look at the Python weakref test suite. I quickly realized
that it appeared to be a richer interface. Maybe that's part of the
problem - the interface we have isn't testable. Take a look here:

http://www.python.org/doc/2.4/lib/module-weakref.html

Then again, maybe I'm way off base. Anyway, I think we should at least
take a look at it and see if we want to steal any ideas from it. The
corresponding test file is test_weakref.py, btw.

Regards,

Dan

···

On Dec 21, 6:50 am, Charles Oliver Nutter <charles.nut...@sun.com> wrote:

Eric Hodel wrote:
> On Dec 20, 2007, at 21:45 PM, Daniel Berger wrote:
>> I follow the same pattern in the tests, but it doesn't raise the
>> expected error. Any ideas?

> Since the GC is conservative, references may exist longer than you
> expect because the item that should be collected is still referenced
> from the stack. I don't know of an easy way around this.

This is probably the largest reason why a weakref suite will be a little
tricky. My best recommendation would be to loop calling GC.start until
the reference gets collected; perhaps with an upper count... "if we
reach 1000 iterations and it's not collected, it isn't working".

6

Yes, the behavior looks very non-deterministic from the outside, it
even depends on naming of the tests. I had better luck making it work
with 1.8 (works as is) but it won't work with 1.9. The first test that
tests the weakref dies does not succeed, and the second does.

This one works for me but it likely depends on exact state of the heap
or something like that which is system specific:

require 'test/unit'
require 'weakref'

class TC_WeakRef < Test::Unit::TestCase
   def setup
      @ref = nil
      @str = "hello"
      GC.enable
   end

   # Fails in 1.9 because WeakRef does not have inspect,
   # and it pretends to be String
   def test_weakref_constructor
      assert_respond_to(WeakRef, :new)
      assert_nothing_raised{ @ref = WeakRef.new(@str) }
      assert_kind_of(WeakRef, @ref)
   end

   # TODO: Figure out why last test fails
   # somehow using the string makes a reference somewhere
   def test_weakref_2
      str = @str.dup

      assert_nothing_raised{ @ref = WeakRef.new(@str) }
      assert_equal(true, @ref.weakref_alive?)
      assert_equal str, @ref

      assert_nothing_raised{ GC.start }
      assert_equal(true, @ref.weakref_alive?)
      assert_equal str, @ref

      assert_nothing_raised{ @str = nil }
      assert_equal(true, @ref.weakref_alive?)
      assert_equal str, @ref

      assert_nothing_raised{ GC.start }
      #assert_equal(false, @ref.weakref_alive?)
      #assert_raise(WeakRef::RefError){ @str = @ref * 3 }
   end

   def test_weakref_is_alive_basic
      assert_nothing_raised{ @ref = WeakRef.new(@str) }
      assert_respond_to(@ref, :weakref_alive?)
   end

   # TODO: Figure out why last test does not fail
   def test_weakref_is_alive
      assert_nothing_raised{ @ref = WeakRef.new(@str) }
      assert_equal(true, @ref.weakref_alive?)

      assert_nothing_raised{ GC.start }
      assert_equal(true, @ref.weakref_alive?)

      assert_nothing_raised{ @str = nil }
      assert_equal(true, @ref.weakref_alive?)

      assert_nothing_raised{ GC.start }
      assert_equal(false, @ref.weakref_alive?)
      assert_raise(WeakRef::RefError){ @str = @ref * 3 }
   end

   def teardown
     @str = nil
      @ref = nil
   end
end

···

On 21/12/2007, Eric Hodel <drbrain@segment7.net> wrote:

On Dec 21, 2007, at 05:50 AM, Charles Oliver Nutter wrote:
> Eric Hodel wrote:
>> On Dec 20, 2007, at 21:45 PM, Daniel Berger wrote:
>>> I follow the same pattern in the tests, but it doesn't raise the
>>> expected error. Any ideas?
>> Since the GC is conservative, references may exist longer than you
>> expect because the item that should be collected is still
>> referenced from the stack. I don't know of an easy way around this.
>
> This is probably the largest reason why a weakref suite will be a
> little tricky. My best recommendation would be to loop calling
> GC.start until the reference gets collected; perhaps with an upper
> count... "if we reach 1000 iterations and it's not collected, it
> isn't working".

If it isn't collected after 1, I doubt it will be collected after
1000. You'd have better luck manipulating the stack (calling methods
recursively that eventually allocate things), but you're still not
guaranteed that any object will be collected when you want with ruby
1.8 or 1.9.

7

Apparently the reason the python weakref is testable is that you can
explicitly delete an object. At least that's what the test code seems
to do. This is not possible in ruby, and the ruby gc is not
deterministic enough to always free all unused objects. So it's not
the weakref what would need changing to make weakref testable but the
memory allocation (and especially deallocation) interface.

Thanks

Michal

···

On 22/12/2007, Daniel Berger <djberg96@gmail.com> wrote:

I took a look at the Python weakref test suite. I quickly realized
that it appeared to be a richer interface. Maybe that's part of the
problem - the interface we have isn't testable. Take a look here:

http://www.python.org/doc/2.4/lib/module-weakref.html

Then again, maybe I'm way off base. Anyway, I think we should at least
take a look at it and see if we want to steal any ideas from it. The
corresponding test file is test_weakref.py, btw.