2011-04-28 14:17:19

by KOSAKI Motohiro

Subject: [PATCH] cpumask: alloc_cpumask_var() use NUMA_NO_NODE

NUMA_NO_NODE and numa_node_id() are different meanings. NUMA_NO_NODE
is obviously recomended fallback.

Signed-off-by: KOSAKI Motohiro <[email protected]>
---
lib/cpumask.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/lib/cpumask.c b/lib/cpumask.c
index 4f6425d..af3e581 100644
--- a/lib/cpumask.c
+++ b/lib/cpumask.c
@@ -131,7 +131,7 @@ EXPORT_SYMBOL(zalloc_cpumask_var_node);
*/
bool alloc_cpumask_var(cpumask_var_t *mask, gfp_t flags)
{
- return alloc_cpumask_var_node(mask, flags, numa_node_id());
+ return alloc_cpumask_var_node(mask, flags, NUMA_NO_NODE);
}
EXPORT_SYMBOL(alloc_cpumask_var);

--
1.7.3.1

2011-05-05 19:46:01

by Andrew Morton

Subject: Re: [PATCH] cpumask: alloc_cpumask_var() use NUMA_NO_NODE

On Thu, 28 Apr 2011 23:17:15 +0900 (JST)
KOSAKI Motohiro <[email protected]> wrote:

> NUMA_NO_NODE and numa_node_id() are different meanings. NUMA_NO_NODE
> is obviously recomended fallback.
>
> Signed-off-by: KOSAKI Motohiro <[email protected]>
> ---
> lib/cpumask.c | 2 +-
> 1 files changed, 1 insertions(+), 1 deletions(-)
>
> diff --git a/lib/cpumask.c b/lib/cpumask.c
> index 4f6425d..af3e581 100644
> --- a/lib/cpumask.c
> +++ b/lib/cpumask.c
> @@ -131,7 +131,7 @@ EXPORT_SYMBOL(zalloc_cpumask_var_node);
> */
> bool alloc_cpumask_var(cpumask_var_t *mask, gfp_t flags)
> {
> - return alloc_cpumask_var_node(mask, flags, numa_node_id());
> + return alloc_cpumask_var_node(mask, flags, NUMA_NO_NODE);
> }
> EXPORT_SYMBOL(alloc_cpumask_var);
>

So effectively this will replace numa_node_id() with numa_mem_id(),
yes? What runtime effects might this have?

2011-05-09 00:29:12

by KOSAKI Motohiro

Subject: Re: [PATCH] cpumask: alloc_cpumask_var() use NUMA_NO_NODE

Hi

> On Thu, 28 Apr 2011 23:17:15 +0900 (JST)
> KOSAKI Motohiro <[email protected]> wrote:
>
> > NUMA_NO_NODE and numa_node_id() are different meanings. NUMA_NO_NODE
> > is obviously recomended fallback.
> >
> > Signed-off-by: KOSAKI Motohiro <[email protected]>
> > ---
> > lib/cpumask.c | 2 +-
> > 1 files changed, 1 insertions(+), 1 deletions(-)
> >
> > diff --git a/lib/cpumask.c b/lib/cpumask.c
> > index 4f6425d..af3e581 100644
> > --- a/lib/cpumask.c
> > +++ b/lib/cpumask.c
> > @@ -131,7 +131,7 @@ EXPORT_SYMBOL(zalloc_cpumask_var_node);
> > */
> > bool alloc_cpumask_var(cpumask_var_t *mask, gfp_t flags)
> > {
> > - return alloc_cpumask_var_node(mask, flags, numa_node_id());
> > + return alloc_cpumask_var_node(mask, flags, NUMA_NO_NODE);
> > }
> > EXPORT_SYMBOL(alloc_cpumask_var);
> >
>
> So effectively this will replace numa_node_id() with numa_mem_id(),
> yes? What runtime effects might this have?

If I understand correctly,

alloc_pages_node(): same effect both NUMA_NO_NODE and numa_node_id()
kmalloc_node(): not same effect NUMA_NO_NODE and numa_node_id()

because

slub.c
---------------------------------------------------
slab_alloc() {
(snip)
if (unlikely(!object || !node_match(c, node)))
// slow path
else {
// fast path

and

static inline int node_match(struct kmem_cache_cpu *c, int node)
{
#ifdef CONFIG_NUMA
if (node != NUMA_NO_NODE && c->node != node)
return 0;
#endif
return 1;
}
---------------------------------------------------
In a nutshell, numa_node_id() reduce slab cache reusing chance.


Oh, I missed slab.c code. It's using numa_mem_id(). thank you correct me!

numa_mem_id() don't have much meanings. It's ia64 HP big machine quirk.
it only affect to improve slab performance a little if users try to allocate
a memory from a cpu within memoryless node . and, 99% users never use such machine.

In a nutshell, NUMA_NO_NODE and numa_node_id() don't have a lot of difference
if users are using SLAB.