Re: [RFC][PATCH 2/2] rcu classic: new algorithm for callbacks-processing(v2)

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

On Wed, Aug 06, 2008 at 03:08:10PM +0800, Lai Jiangshan wrote:
> Paul E. McKenney wrote:
> [...]
> > 
> > Tell me more about percpu_ptr().
> 
> Sorry about this. percpu_ptr is used for dynamic allocation percpu pointer.

Yep, that I knew.

> It seems that we cannot get a pointer from a static declare percpu data
> which can be used as a dynamic allocation percpu data's pointer. 

Sad but true...  Ran into this with SRCU a couple of years back.  :-/

> > 
> [...]
> > 
> > I have a somewhat different goal here.  I want to simplify the memory
> > ordering design without giving up too much performance -- the current
> > state in mainline is much too fragile, in my opinion, especially given
> > that the grace-period code paths are not fastpaths.
> > 
> > Next step -- hierarchical grace-period detection to handle the 4096-CPU
> > machines that I was being buttonholed about at OLS...
> > 
> > Would you be interested in applying your multi-tailed list change to
> > preemptable RCU?
> > 
> It's not necessary. Actually I like one tail per list which is good for
> readability. 
> 
> But in my patch, the most work is combining lists, not
> moving a list to next list, so i use multi-tailed simplify this works
> and others(etc: "if (rdp->nxtlist)" will be changed to be a more
> complex and less readability statement if i use one-tail-per-list)
> 
> These not means multi-tailed is good thing.

It does indeed depend on the details of the implementation.

							Thanx, Paul