Re: [PATCH] x86/mm/cpa: avoid wbinvd() for PREEMPT_RT_FULL

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jan 25, 2017 at 05:12:17PM +0100, Sebastian Andrzej Siewior wrote:
> On 2017-01-21 15:19:15 [+0100], Peter Zijlstra wrote:
> > On Fri, Jan 20, 2017 at 11:42:57PM +0100, John Ogness wrote:
> > > Although wbinvd() is faster than flushing many individual pages, it
> > > blocks the memory bus for "long" periods of time (>100us), thus
> > > directly causing unusually large latencies for PREEMPT_RT_FULL. For
> > > 1024 pages, flushing those pages individually can take up to 2200us,
> > > but the task remains fully preemptible during that time.
> > > 
> > > Cc: Peter Zijlstra <peterz@infradead.org>
> > > Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> > > Signed-off-by: John Ogness <john.ogness@linutronix.de>
> > > ---
> > >  arch/x86/mm/pageattr.c | 5 +++++
> > >  1 file changed, 5 insertions(+)
> > > 
> > > diff --git a/arch/x86/mm/pageattr.c b/arch/x86/mm/pageattr.c
> > > index e3353c9..a182477 100644
> > > --- a/arch/x86/mm/pageattr.c
> > > +++ b/arch/x86/mm/pageattr.c
> > > @@ -214,7 +214,12 @@ static void cpa_flush_array(unsigned long *start, int numpages, int cache,
> > >  			    int in_flags, struct page **pages)
> > >  {
> > >  	unsigned int i, level;
> > > +#ifdef CONFIG_PREEMPT_RT_FULL
> > > +	/* wbinvd() causes ugly latencies, avoid it */
> > > +	unsigned long do_wbinvd = 0;
> > 
> > Arguably we should do the same for CONFIG_PREEMPT and possibly even
> > always, esp. when considering CAT.
> 
> So you want to see this patch again with CONFIG_PREEMPT instead of
> CONFIG_PREEMPT_RT_FULL and also targeting lkml?

Patches should go to LKML irrespective of PREEMPT_RT_FULL or not.

But yes, I think it makes sense for PREEMPT as well.

> I don't get quite the link between wbindv and CAT (assuming it stands
> for Intel's Cache Allocation support). But if you want unconditionally
> want to drop that wbinvd because it is bad for another !RT usecase, fine
> by me :)

Yes, that CAT.

I suspect; but I couldn't find the SDM making any statement what so ever
on this; that when you do wbinvd you kill the _entire_ cache,
irrespective of our CAT mask.

This means that a random task doing wbinvd(), even if it is 'isolated'
with a CAT mask, will affect the performance of all other tasks on that
cache domain.

IOW, your X11 running in some 'system' group will affect your RT
workload on the isolated CPU by killing its cache.