[PATCH 2.6.26-rc3 : schedule] remove unlikely macros in workqueue.c/queue_delayed_work_on

[PATCH 2.6.26-rc3 : schedule] remove unlikely macros in workqueue.c/queue_delayed_work_on

[youquan_song]

cpufreq_ondemand governor call queue_delayed_work_on with entry parameter
"cpu" every sample rate(every logical cpu during 20ms).check the value of
"cpu","cpu>=0" condition meeting rate is over than 90%.

This patch remove the unlikely macros to benefit kernel schedule and
reader comprehension.

Signed-off-by: Youquan Song <youquan.song@intel.com>
---
 workqueue.c    |    2++++++
 1 file changed,     1 insertions(+), 1 deletions(-)

--- linux-2.6/kernel/workqueue.c	2008-05-13 10:10:11.000000000 -0400
+++ linux-2.6-new/kernel/workqueue.c	2008-05-29 09:46:16.000000000 -0400
@@ -230,7 +230,7 @@
		timer->data = (unsigned long)dwork;
		timer->function = delayed_work_timer_fn;

-		if (unlikely(cpu >= 0))
+		if (cpu >= 0)
			add_timer_on(timer, cpu);
		else
			add_timer(timer);

Re: [PATCH 2.6.26-rc3 : schedule] remove unlikely macros in workqueue.c/queue_delayed_work_on

[Andrew Morton]

On Thu, 29 May 2008 00:02:50 -0700 (PDT)
youquan_song@linux.intel.com wrote:

> cpufreq_ondemand governor call queue_delayed_work_on with entry parameter
> "cpu" every sample rate(every logical cpu during 20ms).check the value of
> "cpu","cpu>=0" condition meeting rate is over than 90%.
> 
> This patch remove the unlikely macros to benefit kernel schedule and
> reader comprehension.
> 
> Signed-off-by: Youquan Song <youquan.song@intel.com>
> ---
>  workqueue.c    |    2++++++
>  1 file changed,     1 insertions(+), 1 deletions(-)
> 
> --- linux-2.6/kernel/workqueue.c	2008-05-13 10:10:11.000000000 -0400
> +++ linux-2.6-new/kernel/workqueue.c	2008-05-29 09:46:16.000000000 -0400
> @@ -230,7 +230,7 @@
> 		timer->data = (unsigned long)dwork;
> 		timer->function = delayed_work_timer_fn;
> 
> -		if (unlikely(cpu >= 0))
> +		if (cpu >= 0)
> 			add_timer_on(timer, cpu);
> 		else
> 			add_timer(timer);
> 

This is very much dependent on what workload the machine is running.

There are 189 queue_delayed_work() callsites and they all want the
unlikely() to be there.

There are six queue_delayed_work_on() callsites and they don't want the
unlikely().

Don't know what to do here.  Fortunately it doesn't matter much ;)

-mm has profile-likely-unlikely-macros.patch which can be used to
instrument these things (that feature seems to get broken regularly
though).

But the instrumentation should be performed across a broad range of
workloads.

Re: [PATCH 2.6.26-rc3 : schedule] remove unlikely macros in workqueue.c/queue_delayed_work_on

[Arjan van de Ven]

On Thu, 29 May 2008 13:29:15 -0700
Andrew Morton <akpm@linux-foundation.org> wrote:

> On Thu, 29 May 2008 00:02:50 -0700 (PDT)
> youquan_song@linux.intel.com wrote:
> 
> > cpufreq_ondemand governor call queue_delayed_work_on with entry
> > parameter "cpu" every sample rate(every logical cpu during
> > 20ms).check the value of "cpu","cpu>=0" condition meeting rate is
> > over than 90%.
> > 
> > This patch remove the unlikely macros to benefit kernel schedule and
> > reader comprehension.
> > 
> > Signed-off-by: Youquan Song <youquan.song@intel.com>
> > ---
> >  workqueue.c    |    2++++++
> >  1 file changed,     1 insertions(+), 1 deletions(-)
> > 
> > --- linux-2.6/kernel/workqueue.c	2008-05-13
> > 10:10:11.000000000 -0400 +++
> > linux-2.6-new/kernel/workqueue.c	2008-05-29
> > 09:46:16.000000000 -0400 @@ -230,7 +230,7 @@ timer->data =
> > (unsigned long)dwork; timer->function = delayed_work_timer_fn;
> > 
> > -		if (unlikely(cpu >= 0))
> > +		if (cpu >= 0)
> > 			add_timer_on(timer, cpu);
> > 		else
> > 			add_timer(timer);
> > 
> 
> This is very much dependent on what workload the machine is running.
> 
> There are 189 queue_delayed_work() callsites and they all want the
> unlikely() to be there.
> 
> There are six queue_delayed_work_on() callsites and they don't want
> the unlikely().
> 
> Don't know what to do here.  Fortunately it doesn't matter much ;)
> 
> -mm has profile-likely-unlikely-macros.patch which can be used to
> instrument these things (that feature seems to get broken regularly
> though).
> 
> But the instrumentation should be performed across a broad range of
> workloads.


and on x86 it barely makes any difference at all, at best you can
handwave a little about icache footprint. 

/me ponders checking if we get a code size reduction by defining these
animals to be empty passthrough