Re: [PATCH] It may not be assumed that wake_up(), finish_wait() and co. imply a memory barrier

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

On Fri, Apr 24, 2009 at 06:48:06PM +0100, David Howells wrote:
> Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> 
> > Because there is no memory barrier between #2 and #3, reordering by
> > either the compiler or the CPU might cause the awakener to update the
> > event_indicated flag in #3 -before- completing its update of shared
> > state in #2.
> 
> If the ordering of #2 and #3 is important with respect to each other, then the
> awakener must manually interpolate a barrier of some sort between the two
> _before_ calling wake_up() (or it should wrap them in a lock).
> 
> As I've tried to make clear in my documentation:
> 
> 	Sleeping and waking on an event flagged in global data can be viewed as
> 	an interaction between two pieces of data: ===> the task state of the
> 	task waiting for the event and the global data used to indicate the
> 	event <===.
> 
> the barrier in wake_up() is only concerned with the ordering of #3 vs #6.  That
> is all it _can_ impose an order upon, since #2 and #3 both happen before
> wake_up() is called, and #3 is what causes the sleeper to break out of the
> sleep loop.
> 
> > So, for this to work correctly, don't we need at least an smp_wmb()
> > between #2 and #3 and at least an smp_rmb() between #4 and #5?  And if
> > #2 does reads (but not writes) at least one variable in the shared state
> > that #5 writes to, don't both barriers need to be smp_mb()?
> 
> Yes, but that's beyond the scope of this section.  set_current_state() imposes
> the partial ordering { #1, #4 } and wake_up() the partial ordering { #3, #6 }
> because those are the controlling features of the loop.
> 
> Managing the data beyond that is up to the caller of set_current_state() and
> the caller of wake_up().

Fair enough!

But I would strongly suggest at least a note calling this out, preferably a
"don't do this" example.

							Thanx, Paul