[Click] NotifierQueue on multi-threaded Click problem.

[Eddie Kohler]

Hi Joonwoo,

Joonwoo Park wrote:
> There are some problems.
> The Task::process_pending() calls add_pending() which is reference master's
> member variable internally.
> So to call Task::process_pending(), _task_lock is need to acquired or
> synchronized by another method.

There was a problem here, but not the problem you describe.  If two or more
threads called Master::process_pending() at the same time, the whole locking
structure would break down.  (Duh.)

>> This change is checked in.  Does it work for you?
> 
> I think it has problems too.
> Memorizing my test of the last month, It might _empty_note is need to
> be synchronized.
> IMO the puller and pusher is watching notifier race-conditionally.

It seems that you are basing this on reading code rather than tests.  Can you
confirm that there is a problem in tests?

> I tried a patch and It's doing lock-less way.
> It passes pointers of master's member to process_pending(add_pending)
> to prevent pollution of another thread.
> How do you think?

See above; I think this patch doesn't fix the true problem.  (The failed
assertion was the true problem.)

I have checked in a different patch, though, that is at least getting closer.
 Comment discussion and diff here:

http://www.read.cs.ucla.edu/gitweb?p=click;a=commit;h=ba637c51282065cdae499b9f5cb5320950c11271

Does this work better for you?  Do you understand what it is doing?

Eddie