Re: [PATCH] x86, UV: Fix RTC latency bug by reading replicated cachelines

[Dimitri Sivanich <sivanich@sgi.com>]

On Wed, Jan 20, 2010 at 09:10:20AM +0100, Ingo Molnar wrote:
> 
> * Dimitri Sivanich <sivanich@sgi.com> wrote:
> 
> > +++ linux/drivers/char/uv_mmtimer.c	2010-01-14 11:56:57.000000000 -0600
> > @@ -89,12 +89,19 @@ static long uv_mmtimer_ioctl(struct file
> >  	switch (cmd) {
> >  	case MMTIMER_GETOFFSET:	/* offset of the counter */
> >  		/*
> > -		 * UV RTC register is on its own page
> > +		 * Starting with HUB rev 2.0, the UV RTC register is
> > +		 * replicated across all cachelines of it's own page.
> > +		 * This allows faster simultaneous reads from a given socket.
> > +		 *
> > +		 * The offset returned is in 64 bit units.
> >  		 */
> > -		if (PAGE_SIZE <= (1 << 16))
> > -			ret = ((UV_LOCAL_MMR_BASE | UVH_RTC) & (PAGE_SIZE-1))
> > -				/ 8;
> > -		else
> > +		if (PAGE_SIZE <= (1 << 16)) {
> > +			if (uv_get_min_hub_revision_id() == 1)
> > +				ret = 0;
> > +			else
> > +				ret = ((uv_blade_processor_id() *
> > +					L1_CACHE_BYTES) % PAGE_SIZE) / 8;
> > +		} else
> 
> That 64K PAGE_SIZE check in the ioctl looks rather weird. What is the purpose 
> of it?

Checking for > 64k pages is really a holdover from similar code on ia64.  I've taken it out with a modified version of this patch that I'll send shortly.