[PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Shan Wei]

From: Shan Wei <davidshan@tencent.com>

no change vs v1.

Lots of drivers use this kind to read/write per-cpu variable.
stats = this_cpu_ptr(dp->stats_percpu);
u64_stats_update_begin(&stats->sync);
 		stats->tx_packets++;
u64_stats_update_begin(&stats->sync);


Signed-off-by: Shan Wei <davidshan@tencent.com>
---
 net/openvswitch/datapath.c |    4 ++--
 net/openvswitch/vport.c    |    5 ++---
 2 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/net/openvswitch/datapath.c b/net/openvswitch/datapath.c
index 4c4b62c..77d16a5 100644
--- a/net/openvswitch/datapath.c
+++ b/net/openvswitch/datapath.c
@@ -208,7 +208,7 @@ void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
 	int error;
 	int key_len;
 
-	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
+	stats = this_cpu_ptr(dp->stats_percpu);
 
 	/* Extract flow from 'skb' into 'key'. */
 	error = ovs_flow_extract(skb, p->port_no, &key, &key_len);
@@ -282,7 +282,7 @@ int ovs_dp_upcall(struct datapath *dp, struct sk_buff *skb,
 	return 0;
 
 err:
-	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
+	stats = this_cpu_ptr(dp->stats_percpu);
 
 	u64_stats_update_begin(&stats->sync);
 	stats->n_lost++;
diff --git a/net/openvswitch/vport.c b/net/openvswitch/vport.c
index 03779e8..70af0be 100644
--- a/net/openvswitch/vport.c
+++ b/net/openvswitch/vport.c
@@ -333,8 +333,7 @@ void ovs_vport_receive(struct vport *vport, struct sk_buff *skb)
 {
 	struct vport_percpu_stats *stats;
 
-	stats = per_cpu_ptr(vport->percpu_stats, smp_processor_id());
-
+	stats = this_cpu_ptr(vport->percpu_stats);
 	u64_stats_update_begin(&stats->sync);
 	stats->rx_packets++;
 	stats->rx_bytes += skb->len;
@@ -359,7 +358,7 @@ int ovs_vport_send(struct vport *vport, struct sk_buff *skb)
 	if (likely(sent)) {
 		struct vport_percpu_stats *stats;
 
-		stats = per_cpu_ptr(vport->percpu_stats, smp_processor_id());
+		stats = this_cpu_ptr(vport->percpu_stats);
 
 		u64_stats_update_begin(&stats->sync);
 		stats->tx_packets++;
-- 
1.7.1

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Christoph Lameter]

On Sat, 3 Nov 2012, Shan Wei wrote:

> +++ b/net/openvswitch/datapath.c
> @@ -208,7 +208,7 @@ void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
>  	int error;
>  	int key_len;
>
> -	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
> +	stats = this_cpu_ptr(dp->stats_percpu);
>
>  	/* Extract flow from 'skb' into 'key'. */
>  	error = ovs_flow_extract(skb, p->port_no, &key, &key_len);
> @@ -282,7 +282,7 @@ int ovs_dp_upcall(struct datapath *dp, struct sk_buff *skb,
>  	return 0;
>
>  err:
> -	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
> +	stats = this_cpu_ptr(dp->stats_percpu);
>
>  	u64_stats_update_begin(&stats->sync);
>  	stats->n_lost++;
> diff --git a/net/openvswitch/vport.c b/net/openvswitch/vport.c
> index 03779e8..70af0be 100644
> --- a/net/openvswitch/vport.c
> +++ b/net/openvswitch/vport.c
> @@ -333,8 +333,7 @@ void ovs_vport_receive(struct vport *vport, struct sk_buff *skb)
>  {
>  	struct vport_percpu_stats *stats;
>
> -	stats = per_cpu_ptr(vport->percpu_stats, smp_processor_id());
> -
> +	stats = this_cpu_ptr(vport->percpu_stats);
>  	u64_stats_update_begin(&stats->sync);
>  	stats->rx_packets++;
>  	stats->rx_bytes += skb->len;
> @@ -359,7 +358,7 @@ int ovs_vport_send(struct vport *vport, struct sk_buff *skb)
>  	if (likely(sent)) {
>  		struct vport_percpu_stats *stats;
>
> -		stats = per_cpu_ptr(vport->percpu_stats, smp_processor_id());
> +		stats = this_cpu_ptr(vport->percpu_stats);
>
>  		u64_stats_update_begin(&stats->sync);
>  		stats->tx_packets++;

Use this_cpu_inc(vport->percpu_stats->packets) here?

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Shan Wei]

Christoph Lameter said, at 2012/11/3 1:46:
>>  		u64_stats_update_begin(&stats->sync);
>>  		stats->tx_packets++;
> 
> Use this_cpu_inc(vport->percpu_stats->packets) here?
 
Lots of network drivers use u64_stats_sync infrastructure for statistics
on 32bit or 64bit hosts no matter how many members in per-cpu variable. 

keep them be consistent, so no plan to change them.

Thanks 

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Christoph Lameter]

On Thu, 8 Nov 2012, Shan Wei wrote:

> Christoph Lameter said, at 2012/11/3 1:46:
> >>  		u64_stats_update_begin(&stats->sync);
> >>  		stats->tx_packets++;
> >
> > Use this_cpu_inc(vport->percpu_stats->packets) here?
>
> Lots of network drivers use u64_stats_sync infrastructure for statistics

So they would all have an advantage from the patch.

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Shan Wei]

Christoph Lameter said, at 2012/11/9 1:18:
> On Thu, 8 Nov 2012, Shan Wei wrote:
> 
>> Christoph Lameter said, at 2012/11/3 1:46:
>>>>  		u64_stats_update_begin(&stats->sync);
>>>>  		stats->tx_packets++;
>>>
>>> Use this_cpu_inc(vport->percpu_stats->packets) here?
>>
>> Lots of network drivers use u64_stats_sync infrastructure for statistics
> 
> So they would all have an advantage from the patch.

I will try to do the optimizing next time in another patchset which not included
in this series patchset.

I will submit v3 version of this series after testing today.

Thanks

 
 

[PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Shan Wei]

From: Shan Wei <davidshan@tencent.com>

Signed-off-by: Shan Wei <davidshan@tencent.com>
---
 net/openvswitch/datapath.c |    4 ++--
 net/openvswitch/vport.c    |    5 ++---
 2 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/net/openvswitch/datapath.c b/net/openvswitch/datapath.c
index 4c4b62c..77d16a5 100644
--- a/net/openvswitch/datapath.c
+++ b/net/openvswitch/datapath.c
@@ -208,7 +208,7 @@ void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
 	int error;
 	int key_len;
 
-	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
+	stats = this_cpu_ptr(dp->stats_percpu);
 
 	/* Extract flow from 'skb' into 'key'. */
 	error = ovs_flow_extract(skb, p->port_no, &key, &key_len);
@@ -282,7 +282,7 @@ int ovs_dp_upcall(struct datapath *dp, struct sk_buff *skb,
 	return 0;
 
 err:
-	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
+	stats = this_cpu_ptr(dp->stats_percpu);
 
 	u64_stats_update_begin(&stats->sync);
 	stats->n_lost++;
diff --git a/net/openvswitch/vport.c b/net/openvswitch/vport.c
index 03779e8..70af0be 100644
--- a/net/openvswitch/vport.c
+++ b/net/openvswitch/vport.c
@@ -333,8 +333,7 @@ void ovs_vport_receive(struct vport *vport, struct sk_buff *skb)
 {
 	struct vport_percpu_stats *stats;
 
-	stats = per_cpu_ptr(vport->percpu_stats, smp_processor_id());
-
+	stats = this_cpu_ptr(vport->percpu_stats);
 	u64_stats_update_begin(&stats->sync);
 	stats->rx_packets++;
 	stats->rx_bytes += skb->len;
@@ -359,7 +358,7 @@ int ovs_vport_send(struct vport *vport, struct sk_buff *skb)
 	if (likely(sent)) {
 		struct vport_percpu_stats *stats;
 
-		stats = per_cpu_ptr(vport->percpu_stats, smp_processor_id());
+		stats = this_cpu_ptr(vport->percpu_stats);
 
 		u64_stats_update_begin(&stats->sync);
 		stats->tx_packets++;
-- 
1.7.1

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Christoph Lameter]

On Wed, 31 Oct 2012, Shan Wei wrote:

> --- a/net/openvswitch/datapath.c
> +++ b/net/openvswitch/datapath.c
> @@ -208,7 +208,7 @@ void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
>  	int error;
>  	int key_len;
>
> -	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
> +	stats = this_cpu_ptr(dp->stats_percpu);

Well this is an improvement and may be ok if the preemption is disabled at
this point. There is another possibility here to use this_cpu_read/add/inc
instead of determining the pointer to the local cpu first and then
performing operations on the fields. The pointer relocation with
this_cpu_xxx ops is implicit in the instructions and safe against changing
of processors. It would also save us the determination of a pointer to the
current cpus stats structure.

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Shan Wei]

Christoph Lameter said, at 2012/11/1 1:39:
> On Wed, 31 Oct 2012, Shan Wei wrote:
> 
>> --- a/net/openvswitch/datapath.c
>> +++ b/net/openvswitch/datapath.c
>> @@ -208,7 +208,7 @@ void ovs_dp_process_received_packet(struct vport *p, struct sk_buff *skb)
>>  	int error;
>>  	int key_len;
>>
>> -	stats = per_cpu_ptr(dp->stats_percpu, smp_processor_id());
>> +	stats = this_cpu_ptr(dp->stats_percpu);
> 
> Well this is an improvement and may be ok if the preemption is disabled at
> this point. There is another possibility here to use this_cpu_read/add/inc
> instead of determining the pointer to the local cpu first and then
> performing operations on the fields. The pointer relocation with
> this_cpu_xxx ops is implicit in the instructions and safe against changing
> of processors. It would also save us the determination of a pointer to the
> current cpus stats structure.

yes, this_cpu_ptr just locate the point to current cpu per-cpu data domain.
and then operating [read/write/inc/sub] fields of this per-cpu variable
maybe on other cpu because task is rescheduled for preemption, interrupt.

But for different field in same per-cpu variable, how to guarantee n_missed
and n_hit are from same cpu? 
this_cpu_read(dp->stats_percpu->n_missed);
[processor changed]
this_cpu_read(dp->stats_percpu->n_hit);


In addition, following usage of per_cpu_ptr can be replaced by this_cpu_read.

cpu=get_cpu()
....
*per_cpu_ptr(p,cpu)
....
....
put_cpu()

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Christoph Lameter]

On Thu, 1 Nov 2012, Shan Wei wrote:

> But for different field in same per-cpu variable, how to guarantee n_missed
> and n_hit are from same cpu?
> this_cpu_read(dp->stats_percpu->n_missed);
> [processor changed]
> this_cpu_read(dp->stats_percpu->n_hit);

What does current guarantee that? If it is guaranteed then you can use the
__this_cpu_xxx ops.

> In addition, following usage of per_cpu_ptr can be replaced by this_cpu_read.
>
> cpu=get_cpu()
> ....
> *per_cpu_ptr(p,cpu)
> ....
> ....
> put_cpu()

Right.

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Jesse Gross]

On Thu, Nov 1, 2012 at 7:33 AM, Christoph Lameter <cl@linux.com> wrote:
> On Thu, 1 Nov 2012, Shan Wei wrote:
>
>> But for different field in same per-cpu variable, how to guarantee n_missed
>> and n_hit are from same cpu?
>> this_cpu_read(dp->stats_percpu->n_missed);
>> [processor changed]
>> this_cpu_read(dp->stats_percpu->n_hit);
>
> What does current guarantee that? If it is guaranteed then you can use the
> __this_cpu_xxx ops.

Preemption is disabled in all of the places where writes are done and
all of the reads are from foreign CPUs.

Re: [PATCH 4/9] net: openvswitch: use this_cpu_ptr per-cpu helper

[Christoph Lameter]

On Thu, 1 Nov 2012, Jesse Gross wrote:

> On Thu, Nov 1, 2012 at 7:33 AM, Christoph Lameter <cl@linux.com> wrote:
> > On Thu, 1 Nov 2012, Shan Wei wrote:
> >
> >> But for different field in same per-cpu variable, how to guarantee n_missed
> >> and n_hit are from same cpu?
> >> this_cpu_read(dp->stats_percpu->n_missed);
> >> [processor changed]
> >> this_cpu_read(dp->stats_percpu->n_hit);
> >
> > What does current guarantee that? If it is guaranteed then you can use the
> > __this_cpu_xxx ops.
>
> Preemption is disabled in all of the places where writes are done and
> all of the reads are from foreign CPUs.

Since preemption is disabled no processor change can occur. So its safe to
use __this_cpu ops throughout and they will operate on the current per cpu
area.