sched: current instead rq->current

sched: current instead rq->current

[Colin Vidal]

Hello,

At the beginning of __schedule (kernel/sched/core.c), the current task
is get with rq->curr. I try to to understand why not directly using
current instead?

Since a runqueue is specific to a CPU, it dosen't make sense to get the
the current task of another CPU's runqueue. Yes?

I try the following of Linus's master branch

-       int cpu;
 
-       cpu = smp_processor_id();
-       rq = cpu_rq(cpu);
-       prev = rq->curr;
+       rq = cpu_rq(smp_processor_id());
+       prev = current;

and it seems to work (only tested on x86-64), but... To simple?

Thanks!

Colin

Re: sched: current instead rq->current

[Colin Vidal]

On Mon, 2016-08-15 at 21:30 +0200, Colin Vidal wrote:
> Hello,
> 
> At the beginning of __schedule (kernel/sched/core.c), the current
> task
> is get with rq->curr. I try to to understand why not directly using
> current instead?
> 
> Since a runqueue is specific to a CPU, it dosen't make sense to get
> the
> the current task of another CPU's runqueue. Yes?
> 
> I try the following of Linus's master branch
> 
> -       int cpu;
>  
> -       cpu = smp_processor_id();
> -       rq = cpu_rq(cpu);
> -       prev = rq->curr;
> +       rq = cpu_rq(smp_processor_id());
> +       prev = current;
> 
> and it seems to work (only tested on x86-64), but... To simple?
> 
> Thanks!
> 
> Colin

I realize that I have forgotten to add scheduler subsystem in CC... So
I do it now. Sorry for double post.

Thanks

Colin

Re: sched: current instead rq->current

[Peter Zijlstra]

On Wed, Aug 17, 2016 at 09:39:41PM +0200, Colin Vidal wrote:
> On Mon, 2016-08-15 at 21:30 +0200, Colin Vidal wrote:
> > Hello,
> > 
> > At the beginning of __schedule (kernel/sched/core.c), the current
> > task
> > is get with rq->curr. I try to to understand why not directly using
> > current instead?

> > -       int cpu;
> >  
> > -       cpu = smp_processor_id();
> > -       rq = cpu_rq(cpu);
> > -       prev = rq->curr;
> > +       rq = cpu_rq(smp_processor_id());
> > +       prev = current;
> > 
> > and it seems to work (only tested on x86-64), but... To simple?

its more expensive, rq->curr is a simple dereference (and we need that
cacheline anyway), while current is a weird macro that expands to
potentially a lot of code, depending on the arch.