Yielding a block inside a loop in a thread crashes the VM

[honglilai@…]

Yielding a block inside a loop directly in the Thread block results in a VM crash:

Assertion failed: ((b->flags & flags) == flags), function rb_vm_prepare_block, file dispatcher.cpp, line 1737.
Abort trap

Interestingly, the problem doesn't occur of the loop is wrapped inside a function.

Reproduce with:

require 'thread'

def do_yield
	yield
end

# This doesn't crash.
Thread.new do
	do_yield do
	end
	do_yield do
	end
	do_yield do
	end
end.join

# This doesn't crash either.
def do_yield_loop_in_function
	while true
		do_yield do
		end
	end
end
Thread.new do
	do_yield_loop_in_function
end.join

# This crashes.
Thread.new do
	3.times do
		do_yield do
		end
	end
end.join

# This crashes too.
Thread.new do
	while true
		do_yield do
		end
	end
end.join

[ernest.prabhakar@…]

Cc Me!

[dylan@…]

Cc Me!

[watson1978@…]

In dispatcher.cpp:

When removed flag of VM_BLOCK_THREAD from b->flags in rb_vm_yield_args(),

• Other than a main thread
• returned "*cached = true" in RoxorVM::uncache_or_create_block()

At the above condition, VM_BLOCK_THREAD does not set to b->flags. So, calls yield at second time, assertion failed at "assert((b->flags & flags) == flags);" in rb_vm_prepare_block().

I atach a patch file. Please check it.

[lsansonetti@…]

Thanks for the detective work & patch! It seems clearer now.

I wonder if the following patch isn't safer, though.

Index: dispatcher.cpp
===================================================================
--- dispatcher.cpp	(revision 4369)
+++ dispatcher.cpp	(working copy)
@@ -1338,7 +1338,7 @@
     }
     else {
 	assert(b->dvars_size == dvars_size);
-	assert((b->flags & flags) == flags);
+	assert((b->flags & flags) == (flags & ~VM_BLOCK_THREAD));
     }
 
     b->proc = Qnil;

Actually, I wonder if we shouldn't get rid of this assert.

[watson1978@…]

With your patch,in the following conditions, I think that assertion failed occures.

• b->flags = VM_BLOCK_THREAD
• flags = VM_BLOCK_THREAD
  • b->flags & flags => VM_BLOCK_THREAD
  • (flags & ~VM_BLOCK_THREAD) => 0

[lsansonetti@…]

I see, I think you're right :-) I merged your patch in r4379. Thanks a lot for taking the time to work on this :-)

[watson1978@…]

Thank you for fixing this issue! :)