Re: Thread-unsafe initialization problems in Guile

[Mark H Weaver]

Hi Andy,

Andy Wingo <address@hidden> writes:
> On Thu 29 Nov 2012 23:42, Mark H Weaver <address@hidden> writes:
>
>> SCM
>> scm_local_eval (SCM exp, SCM env)
>> {
>>   static SCM local_eval_var = SCM_BOOL_F;
>>
>>   if (scm_is_false (local_eval_var))
>>     local_eval_var = scm_c_public_variable ("ice-9 local-eval", 
>> "local-eval");
>>
>>   return scm_call_2 (SCM_VARIABLE_REF (local_eval_var), exp, env);
>> }
>>
>> The problem is that it's possible for a thread to see a non-#f value for
>> 'local_eval_var' before it sees the memory it points to properly
>> initialized.
>
> scm_c_public_variable is not idempotent?

It is idempotent as far as I know, but that's not the problem.  The
problem is that in modern weakly-ordered memory models, there is no
guarantee that thread B will see writes in the same order as they were
performed by thread A.

For example, if (ice-9 local-eval) has not yet been loaded when thread A
calls 'scm_local_eval', then thread A will load the (ice-9 local-eval)
module, allocate and initialize the associated SCM variable objects, and
finally write 'local_eval_var' to point to one of these variable
objects.  However, in the absence of thread synchronization, thread B
may see 'local_eval_var' set to a non-#f value before it sees the writes
that initialized the variable object to which 'local_eval_var' points.
In other words, thread B may end up reading a SCM variable object that
has not yet been initialized in its timeline.

For a good introduction to what is needed to write robust multithreaded
code on modern weakly-ordered memory architectures, I recommend the
following article:

http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2480.html

   Regards,
     Mark