Re: [RFC][PATCH 12/13] stop_machine: Remove lglock

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

On Tue, Jun 23, 2015 at 12:55:48PM +0200, Peter Zijlstra wrote:
> On Tue, Jun 23, 2015 at 12:09:32PM +0200, Peter Zijlstra wrote:
> > We can of course slap a percpu-rwsem in, but I wonder if there's
> > anything smarter we can do here.
> 
> Urgh, we cannot use percpu-rwsem here, because that would require
> percpu_down_write_trylock(), and I'm not sure we can get around the
> sync_sched() for that.
> 
> Now try_stop_cpus(), which requires the down_write_trylock() is used to
> implement synchronize_sched_expedited().
> 
> Using sync_sched() to implement sync_sched_expedited would make me
> happy, but it does somewhat defeat the purpose.
> 
> 
> 
> Also, I think _expedited is used too eagerly, look at this:
> 
> +void dm_sync_table(struct mapped_device *md)
> +{
> +       synchronize_srcu(&md->io_barrier);
> +       synchronize_rcu_expedited();
> +}
> 
> sync_srcu() is slow already, why then bother with an
> sync_rcu_expedited() :/

Actually, this code was added in 2013, which was after the new variant of
synchronize_srcu(), which last I checked is reasonably fast in the common
case (no readers and not having tons of concurrent synchronize_srcu()
calls on the same srcu_struct), especially on systems with a small number
of CPUs, courtesy of srcu_read_lock()'s and srcu_read_unlock()'s read-side
memory barriers.

So synchronize_rcu() really would be expected to have quite a bit higher
latency than synchronize_srcu().

							Thanx, Paul