Module#const_get may return incomplete class

[yousuketto]

I've encountered when executing ActiveResource::Base.find_or_create_resource_for.
https://github.com/rails/activeresource/blob/v4.0.0/lib/active_resource/base.rb#L1501-L1520
name method of incomplete class return nil.

simple reproduction

# example.rb
module Xxx
  class << self
    def find_or_create(klass_name)
      self.const_get(klass_name)
    rescue NameError => e
      self.const_set(klass_name, Class.new)
    end
  end
end

ths = []
200.times do |name|
  100.times do |index|
    t = Thread.new(index.to_s, name.to_s) do |i, n|
      klass_name = "Yyy#{n}"
      yyy = Xxx.find_or_create(klass_name)
      unless yyy.name
        print "hit!! #{klass_name}.name == nil\n"
      end
    end
    ths << t
  end
end
ths.each(&:join)

$ jruby -v
jruby 1.7.20 (1.9.3p551) 2015-05-04 3086e6a on Java HotSpot(TM) 64-Bit Server VM 1.8.0_45-b14 +jit [darwin-x86_64]
$ jruby example.rb 
example.rb:6 warning: already initialized constant Yyy6
example.rb:6 warning: already initialized constant Yyy6
example.rb:6 warning: already initialized constant Yyy10
example.rb:6 warning: already initialized constant Yyy11
example.rb:6 warning: already initialized constant Yyy11
...
example.rb:6 warning: already initialized constant Yyy84
example.rb:6 warning: already initialized constant Yyy84
hit!! Yyy85.name == nil
example.rb:6 warning: already initialized constant Yyy92
example.rb:6 warning: already initialized constant Yyy93
...

[kares]

this is all fine on JRuby's end - you should not expect your code to not print warnings since it's not thread-safe ... JRuby simply can not "know" that :

    def find_or_create(klass_name)
      self.const_get(klass_name)
    rescue NameError => e
      self.const_set(klass_name, Class.new)
    end

is intended to be somehow "atomic" ... would report this on Rails as their code assumes MRI (gil)

[yousuketto]

Sorry, I'm weak in English.

I don't think that warning message is a problem.

Yyy85 was executed name method, and returned nil.

hit!! Yyy85.name == nil

Therefore, Yyy85 was incomplete state.
It is a problem that I could get this class of incomplete state.

I think cause that's the order of processing.

1. Class object was registered in the constant table.
   • https://github.com/jruby/jruby/blob/1.7.20/core/src/main/java/org/jruby/RubyModule.java#L3153
2. Constant name was registred.
   • https://github.com/jruby/jruby/blob/1.7.20/core/src/main/java/org/jruby/RubyModule.java#L3159-L3165

[kares]

oh sorry, I missed that completely .. yes the name is an issue (I've actually run into that as well in JI)

[headius]

This is not fixable in JRuby.

MRI determines name in a very peculiar way: when it has not yet been set for a given class, it walks every class in the system accessible via Object and uses that to build the name.

This has some severe performance implications that most people don't know about: https://bugs.ruby-lang.org/issues/11119

Needless to say, we won't be implementing #name that way. Instead, we use the usually-ok logic of assigning name when the class is first set to a constant. This behavior you see in Rails is exploting a very strange edge case, and it won't work under JRuby or Rubinius.

[headius]

We may be able to mitigate this by setting the name before we assign the constant...I think that would be reasonable and be a closer approximation of MRI's behavior.