Re: About lock-less data structure patches

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Apr 07, 2011 at 02:32:06PM -0400, Mathieu Desnoyers wrote:
> * Huang Ying (ying.huang@intel.com) wrote:
> > On 04/06/2011 09:48 AM, Mathieu Desnoyers wrote:
> > > * huang ying (huang.ying.caritas@gmail.com) wrote:
> > [snip]
> > >>>>
> > >>>> OK.  I will change the comments, adding these semantics explanation.
> > >>>> The user should be warned :)
> > >>>
> > >>> Yes, that makes sense. After this generalization step, if you're ok with
> > >>> this, we could aim at moving the implementation from a stack to a queue
> > >>> and provide fifo semantic rather than lifo, so that other users (e.g.
> > >>> call_rcu in the kernel) can start benefiting from it.

Just to be clear...  Currently, call_rcu() works on a per-CPU basis,
so that it can simply disable interrupts and then do the queuing
non-atomically.

However, should it be necessary to cross-queue RCU callbacks in order
to avoid ever executing an RCU callback on a given CPU, then something
like this might become useful.

							Thanx, Paul

> > >> I think that is good to move from stack to queue.
> > >>
> > >> I will send out changed lock-less data structure patchset soon.  And
> > >> we can continue to work on the new lock-less queue at the same time.
> > > 
> > > Sounds like a very good plan! Thanks!
> > 
> > Maybe you can send out your lock-less queue patches, so we can work on that.
> 
> Yep, let's wait until your implementation is finalized and merged, and
> then ping me again so I can cook up a RFC patch turning llist into a
> queue, if it's OK with you.
> 
> Thanks,
> 
> Mathieu
> 
> > 
> > Best Regards,
> > Huang Ying
> 
> -- 
> Mathieu Desnoyers
> Operating System Efficiency R&D Consultant
> EfficiOS Inc.
> http://www.efficios.com