Re: [GIT PULL] RCU changes for v3.4

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

On Fri, Mar 23, 2012 at 02:16:38PM -0700, Paul E. McKenney wrote:
> On Fri, Mar 23, 2012 at 12:39:59PM -0700, Linus Torvalds wrote:
> > On Fri, Mar 23, 2012 at 12:21 PM, Paul E. McKenney
> > <paulmck@linux.vnet.ibm.com> wrote:
> > >>
> > >> Please? Every time I look at some profiles, that silly rcu_read_lock
> > >> is there in the profile. It's annoying. I'd rather see it in the
> > >> function that invokes it.
> > >
> > > Let me see what I can do...
> > 
> > Thanks. To some degree, rcu_read_lock() is the more critical one,
> > because it is often in the much more critical path in the caller. In
> > particular, it's often at the beginning of a function, where a number
> > of arguments are "live", and calling it out-of-line also forces the
> > compiler to then save/restore those arguments (because they are
> > clobbered by the function call).
> > 
> > rcu_read_unlock() is *usually* not as critical, and is obviously much
> > harder to inline anyway due to the whole complexity with needing to
> > check if an RCU sequence has ended. It often is at the end of the
> > function call in the caller, when the only thing like is often just
> > the single return value (if that). So it seldom looks nearly as bad in
> > any profiles, because it doesn't tend to have the same kind of bad
> > impact on the call site.
> 
> Very good to hear!  Especially since I am not seeing how to move
> ->rcu_read_unlock_special to a per-CPU variable given that rcu_boost()
> needs cross-task access to it.  There is probably some obvious trick,
> but I will start with just __rcu_read_lock() for now.

And one obvious trick is a per-CPU pointer to the task-structure variable.
But __rcu_read_lock() first.

							Thanx, Paul