Re: [PATCH 2/2] locking/rwsem: Use a return variable in rwsem_spin_on_owner()

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

On Thu, Apr 09, 2015 at 09:47:36AM -0700, Linus Torvalds wrote:
> On Thu, Apr 9, 2015 at 12:53 AM, Ingo Molnar <mingo@kernel.org> wrote:
> >
> > The point is to generally unify the 'out' paths - i.e. to merge it
> > with the rcu_read_unlock() as well, so that we have really simple
> > gotos and only a single exit path.
> 
> Maybe just have the rcu read-locking be done in the *caller* (possibly
> through using just a helper wrapper function that does nothing but the
> locking), so that you can just do a simple "return false" in the
> function itself.
> 
> That said, it worries me a bit that we do that spinning while holding
> the RCU read lock in the first place. Yes, we stop spinning if
> "need_resched()" is set, but what effect - if any - does all of this
> have on RCU latency? If somebody is waiting for a RCU grace period,
> I'm not seeing that setting need-resched...
> 
> At least with CONFIG_PREEMPT_RCU, the read-unlock is *not* just doing
> a preempt-disable, so it's not necessarily just about need_resched().
> It does all the magic with 'rcu_read_unlock_special.s' too..
> 
> Adding Paul. From a RCU locking standpoint, the thing is basically
> (not the real code, edited down):
> 
>         rcu_read_lock();
>         while (sem->owner == owner) {
>                 if (!owner->on_cpu || need_resched())
>                         break;
>                 cpu_relax_lowlatency();
>         }
>         rcu_read_unlock();
> 
> so we busy-loop while holding the RCU read lock while
> 
>    sem->owner == owner && owner->on_cpu && !need_resched()
> 
> is true. That is usually not very long, but we've already had
> watchdogs go off when we get this wrong, so..
> 
> Paul, comments? Are there particular latency concerns wrt
> CONFIG_PREEMPT_RCU here? Or am I just being silly?

If this was a pure spinlock, then the effects of spinning would overwhelm
any problems from extended grace periods.

But this is a sleeplock.  Of course, we stay in the loop only as long as
the lock holder is actually running.  But given that this is a sleeplock,
I am worried that some lock holders might run for long time periods.
After all, that is one of the traditional uses for a sleeplock.  :-/

If the RCU read-side critical section lasts a few hundred milliseconds,
no problem.  If it lasts for more than 500 milliseconds, I would start
getting concerned.

And if such long-term spins are likely, I cannot resist asking if this
should be instead using SRCU.  If you have your own srcu_struct, you
get to delay your own SRCU grace periods as long as you want.  ;-)

							Thanx, Paul