Bug #15262
Updated by larskanis (Lars Kanis) over 5 years ago
Given the following program: ```ruby require "weakref" Thread.abort_on_exception = true class Adder def self.start_adder(obj, oid) self.start_adder(obj) obj.add end def initialize @qu = Queue.new count @mutex = Mutex.new @mutex.lock begin @count = 10 count.times @count.times do Thread.new(WeakRef.new(self), object_id, &self.class.method(:start_adder)) end count.times do ensure @qu.pop @mutex.unlock end @qu.deq end def add @qu.push @mutex.lock begin @count-=1 if @count == 0 @qu.enq true elsif @count < 0 raise "shouldn't happen" end ensure @mutex.unlock end end end def test_adder 10.times.map do Thread.new do Adder.new end end.each(&:join) end 1000.times 100.times do test_adder end ``` ## Expected behaviour: The program should simply execute without error. This is the case on JRuby but not on MRI. ## Actual behavior: The program stops with a probability of approximately 50% 80% with the following error: ``` $ ruby -W2 adder-test.rb #<Thread:0x0000556e9e06f3f8@adder-test2.rb:6 #<Thread:0x0000562d5089fea8@adder-test.rb:6 run> terminated with exception (report_on_exception is true): Traceback (most recent call last): adder-test2.rb:7:in adder-test.rb:7:in `start_adder': Invalid Reference - probably recycled (WeakRef::RefError) Traceback (most recent call last): adder-test.rb:7:in `start_adder': Invalid Reference - probably recycled (WeakRef::RefError) ``` Although `start_adder` works with a `WeakRef`, the `Adder` object should still be GC marked and therefore kept usable per `WeakRef` until marked, since `Adder.new` exited. This is because `Adder.new` waits for completion of doesn't return before all Threads calls to have called `start_adder`. `start_adder` finished. If `Adder.start_adder` is changed to catch the `WeakRef` error like so: ```ruby def self.start_adder(obj, oid) obj.add rescue WeakRef::RefError end ``` ... then the ruby VM detects a deadlock, which shows that `@qu.pop` is still executed within `initialize`. Therefore `WeakRef#add` should not raise a `WeakRef::RefError` to that point in time, but allow access to the object: ``` Traceback (most recent call last): 5: from adder-test2.rb:35:in `<main>' 4: from adder-test2.rb:35:in `times' 3: from adder-test2.rb:36:in `block in <main>' 2: from adder-test2.rb:32:in `test_adder' 1: from adder-test2.rb:32:in `each' adder-test2.rb:32:in `join': No live threads left. Deadlock? (fatal) 2 threads, 2 sleeps current:0x000056520aa3cee0 main thread:0x000056520a5f2ff0 * #<Thread:0x000056520a644b48 sleep_forever> rb_thread_t:0x000056520a5f2ff0 native:0x00007f7ccf535740 int:0 adder-test2.rb:32:in `join' adder-test2.rb:32:in `each' adder-test2.rb:32:in `test_adder' adder-test2.rb:36:in `block in <main>' adder-test2.rb:35:in `times' adder-test2.rb:35:in `<main>' * #<Thread:0x000056520a913040@adder-test2.rb:29 sleep_forever> rb_thread_t:0x000056520a609fc0 native:0x00007f7ca54d4700 int:0 depended by: tb_thread_id:0x000056520a5f2ff0 adder-test2.rb:18:in `pop' adder-test2.rb:18:in `block in initialize' adder-test2.rb:17:in `times' adder-test2.rb:17:in `initialize' adder-test2.rb:30:in `new' adder-test2.rb:30:in `block (2 levels) in test_adder' ``` I verified this on ruby-trunk, but get mostly the same behavior on all older MRI versions.