Re: [PATCH RFC tip/core/rcu 3/4] rcu: Make RCU_FAST_NO_HZ account for pauses out of idle

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

On Thu, Apr 26, 2012 at 03:00:42PM +0200, Peter Zijlstra wrote:
> On Mon, 2012-04-23 at 09:16 -0700, Paul E. McKenney wrote:
> >  static DEFINE_PER_CPU(int, rcu_dyntick_drain);
> >  static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff);
> >  static DEFINE_PER_CPU(struct timer_list, rcu_idle_gp_timer);
> > +static DEFINE_PER_CPU(unsigned long, rcu_idle_gp_timer_expires);
> > +static DEFINE_PER_CPU(bool, rcu_idle_first_pass);
> > +static DEFINE_PER_CPU(unsigned long, rcu_nonlazy_posted);
> > +static DEFINE_PER_CPU(unsigned long, rcu_nonlazy_posted_snap);
> 
> 
> Wouldn't that all be prettier if it were in a struct of sorts?
> 
> struct rcu_dyntick {
> 	int			drain;
> 	unsigned long		holdoff;
> 	struct timer_list	gp_timer;
> 	unsigned long		gp_timer_expires;
> 	bool			first_pass;
> 	unsigned long		posted;
> 	unsigned long		posted_snap;
> };
> 
> static DEFINE_PER_CPU(struct rcu_dyntick, rcu_dyntick);
> 
> ( fwiw, bool doesn't have a well specified storage type )
> 
> This way you have more control over the placement, variables are forced
> to be together, instead of at the mercy of whatever per_cpu and the
> linker do, and you more clearly see the holes in the structure.
> 
> All the per_cpu() usage should still work like:
> 
>   per_cpu(rcu_dyntick.first_pass, cpu) = 0;

Excellent point -- I have added that transformation to the todo list for 3.6.

I must confess that I had no idea that you could say the above.  I would
have expected to have to say the following:

    per_cpu(rcu_dyntick, cpu).first_pass = 0;

Not much difference either way, though.

							Thanx, Paul