Re: Semantics of smp_mb() [was : Re: [PATCH] Fix RCU race in access of nohz_cpu_mask ]

[Jeremy Higdon <jeremy@sgi.com>]

On Mon, Mar 13, 2006 at 10:39:32AM -0800, Paul E. McKenney wrote:
> On Thu, Dec 15, 2005 at 11:46:26PM -0800, Jeremy Higdon wrote:
> > Roland Dreier got this right.  The purpose of the mmiowb is
> > to ensure that writes to I/O devices while holding a spinlock
> > are ordered with respect to writes issued after the original
> > processor releases and a second processor acquires said
> > spinlock.
> > 
> > A MMIO read would be sufficient, but is much heavier weight.
> > 
> > On the SGI MIPS-based systems, the "sync" instruction was used.
> > On the Altix systems, a register on the hub chip is read.
> > 
> > >From comments by jejb, we're looking at modifying the mmiowb
> > API by adding an argument which would be a register to read
> > from if the architecture in question needs ordering in this
> > way but does not have a lighter weight mechanism like the Altix
> > mmiowb.  Since there will now need to be a width indication,
> > mmiowb will be replaced with mmiowb[bwlq].
> 
> Any progress on this front?  I figured that I would wait to update
> the ordering document until after this change happened, but if it
> is going to be awhile, I should proceed with the current API.
> 
> Thoughts?
> 
> 						Thanx, Paul

Brent Casavant was going to be working on this.  I'll CC him so that
he can indicate the status.

jeremy