Re: [PATCH] SN2 user-MMIO CPU migration

[Robin Holt <holt@sgi.com>]

On Mon, Jan 23, 2006 at 11:04:21PM -0800, Grant Grundler wrote:
> On Tue, Jan 24, 2006 at 12:30:42AM -0600, Brent Casavant wrote:
> > On Mon, 23 Jan 2006, Grant Grundler wrote:
> > > Would an app first need to mmap an uncached address?
> > > Or some other obvious "marker" that might warn the kernel?
> > 
> > I'm not sure an uncached address would be appropriate, as I
> > believe there are other non-IO uncached mappings which can
> > be performed (whatever the fetchop stuff morphed into, for
> > example -- I can't remember what its called now).
> 
> OK - just not many user space programs map uncache IO space.

The mspec driver does.  It has not gotten into the mainline kernel yet,
but we continue to build an add-on module for the necessary releases.
To prevent this out of order write from happening on those machines,
we require the process to be pinned to a single cpu.  For MPI, this
is a normal way to operate for other reasons as well.

Brent, if you find a way to prevent this out-of-order write, please
remind me to include it in the mspec driver.

Thanks,
Robin