Subj : Re: experimental smr-based 100% lock-free eventcount schema for windows...
To   : comp.programming.threads
From : Joseph Seigh
Date : Wed Feb 02 2005 07:52 am

On Tue, 1 Feb 2005 18:19:40 -0800, SenderX <xxx@xxx.com> wrote:

> There a race-condition wrt incrementing the eventcount and signalling...
>
>
>> typedef struct ac_event_count_
>> {
>> volatile ac_waitset_t *active_waitset;
>> DWORD ec;
>>
>> } ac_event_count_t;
>
> needs to be modified atomically using dwcas. Here is what I am getting at
> for the waiting side:


I don't think you have to atomically do both.  For signaling you can

	1) atomically increment the eventcount
	2) atomically swap the current event object with a new
	   unsignaled event object.
	3) signal the swapped out (old) event object and
          drop the reference to it.  Old event object will be
	   GC'd when there are no waiters using it.

and waiting is

	1) get pointer to current event object.
	2) atomically compare current eventcount with local copy to
	   see if eventcount has changed
	3) if eventcount unchanged, wait on event object.

The pointers used for the event object are some form of smart pointer,
e.g. atomic_ptr, rcu, smr, etc...


-- 
Joe Seigh

.