Re: [PATCH 6/7] sched, x86: Provide a per-cpu preempt_count implementation

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Sep 10, 2013 at 07:02:51AM -0700, Eric Dumazet wrote:
> On Tue, 2013-09-10 at 15:08 +0200, Peter Zijlstra wrote:
> 
> > +static __always_inline int preempt_count(void)
> > +{
> > +	return __this_cpu_read_4(__preempt_count) & ~PREEMPT_NEED_RESCHED;
> > +}
> 
> Not sure why you used the _4 prefix on all accessors ?

Last time I tried using the proper this_cpu* stuff that all exploded in
my face due to header recursion hell, so I've limited myself to what's
available in arch/x86/include/asm/percpu.h.

It was a few weeks ago though and maybe I just didn't try hard enough.

> > +#ifdef CONFIG_PREEMPT_COUNT
> > +	/*
> > +	 * If it were not for PREEMPT_ACTIVE we could guarantee that the
> > +	 * preempt_count of all tasks was equal here and this would not be
> > +	 * needed.
> > +	 */
> > +	task_thread_info(prev_p)->saved_preempt_count = __raw_get_cpu_var(__preempt_count);
> 
> 	this_cpu_read(__preempt_count) ?
> 
> > +	__raw_get_cpu_var(__preempt_count) = task_thread_info(next_p)->saved_preempt_count;
> 
> 	this_cpu_write(__preempt_count,
>                        task_thread_info(next_p)->saved_preempt_count;

OK, that does indeed generate slightly better code.