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

[Peter Zijlstra <peterz@infradead.org>]

Patches really should go to lkml, if you want to bug a user list with
them I won't stop you, but please don't hide actual useful content.

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.