percpu data for non-existant cpus

percpu data for non-existant cpus

[Jack Steiner]

It was neccessary in 2.4 to allocate cpu_data areas for
non-existent cpus. I forget where, but some generic
initialization code expected to be able to reference cpu_data 
for all cpus up to NR_CPUS.

Does 2.6 have the same requirement or can I allocate cpu_data
areas only for cpus that exist or might exist (cpu_possible()).


-- 
Thanks

Jack Steiner (steiner@sgi.com)          651-683-5302
Principal Engineer                      SGI - Silicon Graphics, Inc.

Re: percpu data for non-existant cpus

[Takayoshi Kochi]

[-- Attachment #1: Type: Text/Plain, Size: 859 bytes --]

Hi,

From: Jack Steiner <steiner@sgi.com>
Subject: percpu data for non-existant cpus
Date: Mon, 3 Nov 2003 18:55:24 -0600

> It was neccessary in 2.4 to allocate cpu_data areas for
> non-existent cpus. I forget where, but some generic
> initialization code expected to be able to reference cpu_data 
> for all cpus up to NR_CPUS.
> 
> Does 2.6 have the same requirement or can I allocate cpu_data
> areas only for cpus that exist or might exist (cpu_possible()).

The current discontig (for both 2.4 & 2.6) only allocate
cpu_data on existing nodes (which means existing cpus
eventually), doesn't it?

We've found wrap_mmu_context() refers to all percpu data
and if the any of the cpu_data is not valid,
it may cause panic.

Attached patches are for 2.4 & 2.6, which are harmless
even when percpu data are allocated for non-existing
cpus.

---
Takayoshi Kochi

[-- Attachment #2: wrap_mmu_context-2.4.patch --]
[-- Type: Text/Plain, Size: 395 bytes --]

--- linux-2.4.22/arch/ia64/mm/tlb.c.orig	Fri Sep 12 09:45:18 2003
+++ linux-2.4.22/arch/ia64/mm/tlb.c	Wed Oct  8 17:56:03 2003
@@ -75,7 +75,7 @@
 	 * and because interrupts are disabled during context switch.
 	 */
 	for (i = 0; i < NR_CPUS; ++i)
-		if (i != smp_processor_id())
+		if (cpu_online(i) && (i != smp_processor_id()))
 			cpu_data(i)->need_tlb_flush = 1;
 	local_flush_tlb_all();
 }

[-- Attachment #3: wrap_mmu_context-2.6.patch --]
[-- Type: Text/Plain, Size: 366 bytes --]

--- linux-2.6.0-test6.orig/arch/ia64/mm/tlb.c	Sun Sep 28 09:50:38 2003
+++ linux-2.6.0-test6/arch/ia64/mm/tlb.c	Wed Oct  8 17:58:42 2003
@@ -77,7 +77,7 @@
 	{
 		int cpu = get_cpu(); /* prevent preemption/migration */
 		for (i = 0; i < NR_CPUS; ++i)
-			if (i != cpu)
+			if (cpu_online(i) && (i != cpu))
 				per_cpu(ia64_need_tlb_flush, i) = 1;
 		put_cpu();
 	}

Re: percpu data for non-existant cpus

[Matthew Wilcox]

On Tue, Nov 04, 2003 at 10:43:49AM +0900, Takayoshi Kochi wrote:
> --- linux-2.6.0-test6.orig/arch/ia64/mm/tlb.c	Sun Sep 28 09:50:38 2003
> +++ linux-2.6.0-test6/arch/ia64/mm/tlb.c	Wed Oct  8 17:58:42 2003
> @@ -77,7 +77,7 @@
>  	{
>  		int cpu = get_cpu(); /* prevent preemption/migration */
>  		for (i = 0; i < NR_CPUS; ++i)
> -			if (i != cpu)
> +			if (cpu_online(i) && (i != cpu))
>  				per_cpu(ia64_need_tlb_flush, i) = 1;
>  		put_cpu();
>  	}

Surely this should rather be...

	{
		int cpu = get_cpu(); /* prevent preemption/migration */
-		for (i = 0; i < NR_CPUS; ++i)
+		for_each_cpu(i, cpu_online_map)
			if (i != cpu)
				per_cpu(ia64_need_tlb_flush, i) = 1;
		put_cpu();

I suppose one could create a new cpu_map based on cpu_online_map &~ cpu
but that feels rather more heavyweight than a test inside a loop.

-- 
"It's not Hollywood.  War is real, war is primarily not about defeat or
victory, it is about death.  I've seen thousands and thousands of dead bodies.
Do you think I want to have an academic debate on this subject?" -- Robert Fisk