[PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[Shan Wei]

From: Shan Wei <davidshan@tencent.com>

Signed-off-by: Shan Wei <davidshan@tencent.com>
---
no changes vs v2.
---
 kernel/rcutree.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/kernel/rcutree.c b/kernel/rcutree.c
index 74df86b..441b945 100644
--- a/kernel/rcutree.c
+++ b/kernel/rcutree.c
@@ -1960,7 +1960,7 @@ static void force_quiescent_state(struct rcu_state *rsp)
 	struct rcu_node *rnp_old = NULL;
 
 	/* Funnel through hierarchy to reduce memory contention. */
-	rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
+	rnp = __this_cpu_read(rsp->rda->mynode);
 	for (; rnp != NULL; rnp = rnp->parent) {
 		ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
 		      !raw_spin_trylock(&rnp->fqslock);
-- 
1.7.1

Re: [PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[Christoph Lameter]

On Fri, 9 Nov 2012, Shan Wei wrote:

>  	/* Funnel through hierarchy to reduce memory contention. */
> -	rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
> +	rnp = __this_cpu_read(rsp->rda->mynode);
>  	for (; rnp != NULL; rnp = rnp->parent) {

Acked-by: Christoph Lameter <cl@linux.com>

[PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[Shan Wei]

From: Shan Wei <davidshan@tencent.com>

Signed-off-by: Shan Wei <davidshan@tencent.com>
Acked-by: Christoph Lameter <cl@linux.com>
---
no changes vs v3,v2.
---
 kernel/rcutree.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/kernel/rcutree.c b/kernel/rcutree.c
index 74df86b..441b945 100644
--- a/kernel/rcutree.c
+++ b/kernel/rcutree.c
@@ -1960,7 +1960,7 @@ static void force_quiescent_state(struct rcu_state *rsp)
 	struct rcu_node *rnp_old = NULL;
 
 	/* Funnel through hierarchy to reduce memory contention. */
-	rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
+	rnp = __this_cpu_read(rsp->rda->mynode);
 	for (; rnp != NULL; rnp = rnp->parent) {
 		ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
 		      !raw_spin_trylock(&rnp->fqslock);
-- 
1.7.1

Re: [PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[Shan Wei]

Shan Wei said, at 2012/11/13 9:52:
> From: Shan Wei <davidshan@tencent.com>
> 
> Signed-off-by: Shan Wei <davidshan@tencent.com>
> Acked-by: Christoph Lameter <cl@linux.com>

Paul, would you like to pick it up to your tree?

> ---
> no changes vs v3,v2.
> ---
>  kernel/rcutree.c |    2 +-
>  1 files changed, 1 insertions(+), 1 deletions(-)
> 
> diff --git a/kernel/rcutree.c b/kernel/rcutree.c
> index 74df86b..441b945 100644
> --- a/kernel/rcutree.c
> +++ b/kernel/rcutree.c
> @@ -1960,7 +1960,7 @@ static void force_quiescent_state(struct rcu_state *rsp)
>  	struct rcu_node *rnp_old = NULL;
>  
>  	/* Funnel through hierarchy to reduce memory contention. */
> -	rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
> +	rnp = __this_cpu_read(rsp->rda->mynode);
>  	for (; rnp != NULL; rnp = rnp->parent) {
>  		ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
>  		      !raw_spin_trylock(&rnp->fqslock);
> 

Re: [PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[Paul E. McKenney]

On Fri, Nov 16, 2012 at 04:32:50PM +0800, Shan Wei wrote:
> Shan Wei said, at 2012/11/13 9:52:
> > From: Shan Wei <davidshan@tencent.com>
> > 
> > Signed-off-by: Shan Wei <davidshan@tencent.com>
> > Acked-by: Christoph Lameter <cl@linux.com>
> 
> Paul, would you like to pick it up to your tree?

Hello, Shan Wei,

If you either show me some significant performance benefits or get me
an independent Tested-by, in both cases on a range of hardware (e.g.,
x86 on the one hand and ARM or Power on the other), then I will queue it.

I wasn't prioritizing this one very high because it does not appear
to be on any sort of fastpath.  If I am wrong about that, then you
have a good performance-benefit case, right?  ;-)

							Thanx, Paul

> > ---
> > no changes vs v3,v2.
> > ---
> >  kernel/rcutree.c |    2 +-
> >  1 files changed, 1 insertions(+), 1 deletions(-)
> > 
> > diff --git a/kernel/rcutree.c b/kernel/rcutree.c
> > index 74df86b..441b945 100644
> > --- a/kernel/rcutree.c
> > +++ b/kernel/rcutree.c
> > @@ -1960,7 +1960,7 @@ static void force_quiescent_state(struct rcu_state *rsp)
> >  	struct rcu_node *rnp_old = NULL;
> >  
> >  	/* Funnel through hierarchy to reduce memory contention. */
> > -	rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
> > +	rnp = __this_cpu_read(rsp->rda->mynode);
> >  	for (; rnp != NULL; rnp = rnp->parent) {
> >  		ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
> >  		      !raw_spin_trylock(&rnp->fqslock);
> > 
> 

Re: [PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[Christoph Lameter]

On Fri, 16 Nov 2012, Paul E. McKenney wrote:

> If you either show me some significant performance benefits or get me
> an independent Tested-by, in both cases on a range of hardware (e.g.,
> x86 on the one hand and ARM or Power on the other), then I will queue it.

Just putting the code generated for x86 before and after side
by side would be enough to convince you I think.

> I wasn't prioritizing this one very high because it does not appear
> to be on any sort of fastpath.  If I am wrong about that, then you
> have a good performance-benefit case, right?  ;-)

I do not think this needs to be a priority item. Just stick it in the tree
somewhere to merge for the next merge period.

Re: [PATCH v3 6/9] rcu: use __this_cpu_read helper instead of per_cpu_ptr(p, raw_smp_processor_id())

[Paul E. McKenney]

On Fri, Nov 16, 2012 at 05:42:14PM +0000, Christoph Lameter wrote:
> 
> On Fri, 16 Nov 2012, Paul E. McKenney wrote:
> 
> > If you either show me some significant performance benefits or get me
> > an independent Tested-by, in both cases on a range of hardware (e.g.,
> > x86 on the one hand and ARM or Power on the other), then I will queue it.
> 
> Just putting the code generated for x86 before and after side
> by side would be enough to convince you I think.

If accompanied by similar before/after code for ARM or Power, sure.

							Thanx, Paul

> > I wasn't prioritizing this one very high because it does not appear
> > to be on any sort of fastpath.  If I am wrong about that, then you
> > have a good performance-benefit case, right?  ;-)
> 
> I do not think this needs to be a priority item. Just stick it in the tree
> somewhere to merge for the next merge period.
>