[PATCH] - Speed up boot - skip unnecessary clock calibration

[PATCH] - Speed up boot - skip unnecessary clock calibration

[Jack Steiner]

Skip clock calibration if cpu being brought online is exactly the same
speed, stepping, etc., as the previous cpu. This significantly reduces
the time to boot very large systems. 

	Signed-off-by: Jack Steiner <steiner@sgi.com>


Index: linux/arch/ia64/kernel/smpboot.c
=================================--- linux.orig/arch/ia64/kernel/smpboot.c	2007-03-26 15:04:29.322256475 -0500
+++ linux/arch/ia64/kernel/smpboot.c	2007-03-27 08:26:41.914042190 -0500
@@ -424,7 +424,19 @@ smp_callin (void)
 	 * Get our bogomips.
 	 */
 	ia64_init_itm();
-	calibrate_delay();
+
+	/*
+	 * Delay calibration can be skipped if new processor is identical to the
+	 * previous processor.
+	 */
+	if (local_cpu_data->itc_freq != per_cpu(cpu_info, cpuid - 1).itc_freq ||
+			local_cpu_data->proc_freq != per_cpu(cpu_info, cpuid - 1).proc_freq ||
+	    		local_cpu_data->features != per_cpu(cpu_info, cpuid - 1).features ||
+	    		local_cpu_data->revision != per_cpu(cpu_info, cpuid - 1).revision ||
+	    		local_cpu_data->family != per_cpu(cpu_info, cpuid - 1).family ||
+	    		local_cpu_data->archrev != per_cpu(cpu_info, cpuid - 1).archrev ||
+	    		local_cpu_data->model != per_cpu(cpu_info, cpuid - 1).model)
+		calibrate_delay();
 	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
 
 #ifdef CONFIG_IA32_SUPPORT

Re: [PATCH] - Speed up boot - skip unnecessary clock calibration

[Matthew Wilcox]

On Tue, Mar 27, 2007 at 08:29:54AM -0500, Jack Steiner wrote:
> +	if (local_cpu_data->itc_freq != per_cpu(cpu_info, cpuid - 1).itc_freq ||
> +			local_cpu_data->proc_freq != per_cpu(cpu_info, cpuid - 1).proc_freq ||
> +	    		local_cpu_data->features != per_cpu(cpu_info, cpuid - 1).features ||
> +	    		local_cpu_data->revision != per_cpu(cpu_info, cpuid - 1).revision ||
> +	    		local_cpu_data->family != per_cpu(cpu_info, cpuid - 1).family ||
> +	    		local_cpu_data->archrev != per_cpu(cpu_info, cpuid - 1).archrev ||
> +	    		local_cpu_data->model != per_cpu(cpu_info, cpuid - 1).model)
> +		calibrate_delay();

If you use a temporary variable, you improve speed a little, and you
don't exceed 80 columns:

+	struct cpuinfo_ia64 *last_cpu_info = cpu_data(cpuid - 1);
+	struct cpuinfo_ia64 *this_cpu_info = local_cpu_data;
+	if (last_cpu_info->itc_freq != this_cpu_info->itc_freq ||
+	    last_cpu_info->proc_freq != this_cpu_info->proc_freq ||
+	    last_cpu_info->features != this_cpu_info->features ||
+	    last_cpu_info->revision != this_cpu_info->revision ||
+	    last_cpu_info->family != this_cpu_info->family ||
+	    last_cpu_info->archrev != this_cpu_info->archrev ||
+	    last_cpu_info->model != this_cpu_info->model)
+		calibrate_delay();

RE: [PATCH] - Speed up boot - skip unnecessary clock calibration

[Luck, Tony]

> Skip clock calibration if cpu being brought online is exactly the same
> speed, stepping, etc., as the previous cpu. This significantly reduces
> the time to boot very large systems. 

Do we scan through the cpus in a way that makes this really useful I.e.
pounding through the cpus on a node - which are likely to be the
same - before going on to the next node?  Just looking at the
previous cpu won't help as much if we scan "breadth-first".

-Tony

RE: [PATCH] - Speed up boot - skip unnecessary clock calibration

[Yu, Fenghua]

>+	 */
>+	if (local_cpu_data->itc_freq != per_cpu(cpu_info, cpuid -
1).itc_freq >||
>+			local_cpu_data->proc_freq != per_cpu(cpu_info,
cpuid - >1).proc_freq ||

BSP (cupid=0) should always calibrate delay. Otherwise the patch won't
work.

if (cupid=0)
 calibrate_delay()
else
 your patch.

Thanks.

-Fenghua

Re: [PATCH] - Speed up boot - skip unnecessary clock calibration

[Jack Steiner]

On Tue, Mar 27, 2007 at 12:17:17PM -0700, Yu, Fenghua wrote:
> >+	 */
> >+	if (local_cpu_data->itc_freq != per_cpu(cpu_info, cpuid -
> 1).itc_freq >||
> >+			local_cpu_data->proc_freq != per_cpu(cpu_info,
> cpuid - >1).proc_freq ||
> 
> BSP (cupid=0) should always calibrate delay. Otherwise the patch won't
> work.

True, but the call to calibrate the delay on cpu 0 comes from a different
place. See the call in init/main.c.

The code in smp_callin() is only called for cpus 1, 2, ...


-- jack

Re: [PATCH] - Speed up boot - skip unnecessary clock calibration

[Jack Steiner]

Skip clock calibration if cpu being brought online is exactly the same
speed, stepping, etc., as the previous cpu. This significantly reduces
the time to boot very large systems. 

	Signed-off-by: Jack Steiner <steiner@sgi.com>

----
Add suggestion from Matthew Wilcox to use local variables to shorten lines.




Index: linux/arch/ia64/kernel/smpboot.c
=================================--- linux.orig/arch/ia64/kernel/smpboot.c	2007-03-27 10:08:13.664475719 -0500
+++ linux/arch/ia64/kernel/smpboot.c	2007-03-27 14:20:44.070836177 -0500
@@ -375,6 +375,7 @@ static void __devinit
 smp_callin (void)
 {
 	int cpuid, phys_id, itc_master;
+	struct cpuinfo_ia64 *last_cpuinfo, *this_cpuinfo;
 	extern void ia64_init_itm(void);
 	extern volatile int time_keeper_id;
 
@@ -424,7 +425,21 @@ smp_callin (void)
 	 * Get our bogomips.
 	 */
 	ia64_init_itm();
-	calibrate_delay();
+
+	/*
+	 * Delay calibration can be skipped if new processor is identical to the
+	 * previous processor.
+	 */
+	last_cpuinfo = cpu_data(cpuid - 1);
+	this_cpuinfo = local_cpu_data;
+	if (last_cpuinfo->itc_freq != this_cpuinfo->itc_freq ||
+	    last_cpuinfo->proc_freq != this_cpuinfo->proc_freq ||
+	    last_cpuinfo->features != this_cpuinfo->features ||
+	    last_cpuinfo->revision != this_cpuinfo->revision ||
+	    last_cpuinfo->family != this_cpuinfo->family ||
+	    last_cpuinfo->archrev != this_cpuinfo->archrev ||
+	    last_cpuinfo->model != this_cpuinfo->model)
+		calibrate_delay();
 	local_cpu_data->loops_per_jiffy = loops_per_jiffy;
 
 #ifdef CONFIG_IA32_SUPPORT