Re: [rfc patch] x86,uv: -rt conversions

[Mike Galbraith <bitbucket@online.de>]

On Fri, 2013-06-21 at 15:52 +0200, Sebastian Andrzej Siewior wrote: 
> * Mike Galbraith | 2013-06-19 16:26:18 [+0200]:
> 
> >
> >The below is what I use to run -rt on UV boxen, diff generated against
> >3.8-rt.  Should the lock conversions perhaps wander to mainline?
> 
> They seem to come from NMI or irq off region so if you can't change this
> why not.

Maybe SGI can, they know what a bau is.  I only made their locks happy.

> >Signed-off-by: Mike Galbraith <bitbucket@online.de>
> >
> >Index: linux-2.6/arch/x86/platform/uv/uv_time.c
> >===================================================================
> >--- linux-2.6.orig/arch/x86/platform/uv/uv_time.c
> >+++ linux-2.6/arch/x86/platform/uv/uv_time.c
> >@@ -300,13 +300,18 @@ static int uv_rtc_unset_timer(int cpu, i
> > static cycle_t uv_read_rtc(struct clocksource *cs)
> > {
> > 	unsigned long offset;
> >+	cycle_t cycles;
> > 
> >+	migrate_disable();
> > 	if (uv_get_min_hub_revision_id() == 1)
> > 		offset = 0;
> > 	else
> > 		offset = (uv_blade_processor_id() * L1_CACHE_BYTES) % PAGE_SIZE;
> > 
> >-	return (cycle_t)uv_read_local_mmr(UVH_RTC | offset);
> >+	cycles = (cycle_t)uv_read_local_mmr(UVH_RTC | offset);
> >+	migrate_enable();
> >+
> >+	return cycles;
> > }

> You try to ensure not to switch CPUs between uv_blade_processor_id() and
> uv_read_local_mmr()'s final HW access, right?

Yeah.  I originally did preempt_disable/enable() a couple years ago
while convincing a UV to boot/run 2.6.33-rt (x2apic dmar intr_remap
etc), it turned into migrate_disable/enable() as time passed.  I'm not
so sure it's really enough, but I was recently given a UV2000 to tinker
with, and both it and old UV100 box boot and (seem to) run fine as is.

-Mike