[PATCH net-next-2.6] rps: Various optimizations

[PATCH net-next-2.6] rps: Various optimizations

[Eric Dumazet]

Introduce ____napi_schedule() helper for callers in irq disabled
contexts. rps_trigger_softirq() becomes a leaf function.

Use container_of() in process_backlog() instead of accessing per_cpu
address.

Use a custom inlined version of __napi_complete() in process_backlog()
to avoid one locked instruction :

 only current cpu owns and manipulates this napi,
 and NAPI_STATE_SCHED is the only possible flag set on backlog.
 we can use a plain write instead of clear_bit(),
 and we dont need an smp_mb() memory barrier, since RPS is on,
 backlog is protected by a spinlock.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
---
 net/core/dev.c |   28 ++++++++++++++++++++++------
 1 files changed, 22 insertions(+), 6 deletions(-)

diff --git a/net/core/dev.c b/net/core/dev.c
index 36d53be..c6861e4 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -2205,6 +2205,14 @@ int netdev_max_backlog __read_mostly = 1000;
 int netdev_budget __read_mostly = 300;
 int weight_p __read_mostly = 64;            /* old backlog weight */
 
+/* Called with irq disabled */
+static inline void ____napi_schedule(struct softnet_data *sd,
+				     struct napi_struct *napi)
+{
+	list_add_tail(&napi->poll_list, &sd->poll_list);
+	__raise_softirq_irqoff(NET_RX_SOFTIRQ);
+}
+
 #ifdef CONFIG_RPS
 
 /* One global table that all flow-based protocols share. */
@@ -2363,7 +2371,7 @@ static void rps_trigger_softirq(void *data)
 {
 	struct softnet_data *sd = data;
 
-	__napi_schedule(&sd->backlog);
+	____napi_schedule(sd, &sd->backlog);
 	sd->received_rps++;
 }
 
@@ -2421,7 +2429,7 @@ enqueue:
 		/* Schedule NAPI for backlog device */
 		if (napi_schedule_prep(&sd->backlog)) {
 			if (!rps_ipi_queued(sd))
-				__napi_schedule(&sd->backlog);
+				____napi_schedule(sd, &sd->backlog);
 		}
 		goto enqueue;
 	}
@@ -3280,7 +3288,7 @@ static void net_rps_action_and_irq_enable(struct softnet_data *sd)
 static int process_backlog(struct napi_struct *napi, int quota)
 {
 	int work = 0;
-	struct softnet_data *sd = &__get_cpu_var(softnet_data);
+	struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
 
 #ifdef CONFIG_RPS
 	/* Check if we have pending ipi, its better to send them now,
@@ -3313,7 +3321,16 @@ static int process_backlog(struct napi_struct *napi, int quota)
 						   &sd->process_queue);
 		}
 		if (qlen < quota - work) {
-			__napi_complete(napi);
+			/*
+			 * Inline a custom version of __napi_complete().
+			 * only current cpu owns and manipulates this napi,
+			 * and NAPI_STATE_SCHED is the only possible flag set on backlog.
+			 * we can use a plain write instead of clear_bit(),
+			 * and we dont need an smp_mb() memory barrier.
+			 */
+			list_del(&napi->poll_list);
+			napi->state = 0;
+
 			quota = work + qlen;
 		}
 		rps_unlock(sd);
@@ -3334,8 +3351,7 @@ void __napi_schedule(struct napi_struct *n)
 	unsigned long flags;
 
 	local_irq_save(flags);
-	list_add_tail(&n->poll_list, &__get_cpu_var(softnet_data).poll_list);
-	__raise_softirq_irqoff(NET_RX_SOFTIRQ);
+	____napi_schedule(&__get_cpu_var(softnet_data), n);
 	local_irq_restore(flags);
 }
 EXPORT_SYMBOL(__napi_schedule);

Re: [PATCH net-next-2.6] rps: Various optimizations

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Thu, 06 May 2010 10:58:42 +0200

> Introduce ____napi_schedule() helper for callers in irq disabled
> contexts. rps_trigger_softirq() becomes a leaf function.
> 
> Use container_of() in process_backlog() instead of accessing per_cpu
> address.
> 
> Use a custom inlined version of __napi_complete() in process_backlog()
> to avoid one locked instruction :
> 
>  only current cpu owns and manipulates this napi,
>  and NAPI_STATE_SCHED is the only possible flag set on backlog.
>  we can use a plain write instead of clear_bit(),
>  and we dont need an smp_mb() memory barrier, since RPS is on,
>  backlog is protected by a spinlock.
> 
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>

Looks great, applied, thanks Eric.

Re: [PATCH net-next-2.6] rps: Various optimizations

[Eric Dumazet]

Le jeudi 06 mai 2010 à 22:07 -0700, David Miller a écrit :

> Looks great, applied, thanks Eric.

Thanks, I have a followup to avoid one atomic in enqueue phase too ;)

[PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog

[Eric Dumazet]

Le vendredi 07 mai 2010 à 07:16 +0200, Eric Dumazet a écrit :
> Le jeudi 06 mai 2010 à 22:07 -0700, David Miller a écrit :
> 
> > Looks great, applied, thanks Eric.
> 
> Thanks, I have a followup to avoid one atomic in enqueue phase too ;)
> 

[PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog

If CONFIG_SMP=y, then we own a queue spinlock, we can avoid the atomic
test_and_set_bit() from napi_schedule_prep().

We now have same number of atomic ops per netif_rx() calls than with
pre-RPS kernel.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
---
diff --git a/net/core/dev.c b/net/core/dev.c
index 32611c8..49fa5a6 100644
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -2426,8 +2426,10 @@ enqueue:
 			return NET_RX_SUCCESS;
 		}
 
-		/* Schedule NAPI for backlog device */
-		if (napi_schedule_prep(&sd->backlog)) {
+		/* Schedule NAPI for backlog device
+		 * We can use non atomic operation since we own the queue lock
+		 */
+		if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
 			if (!rps_ipi_queued(sd))
 				____napi_schedule(sd, &sd->backlog);
 		}

Re: [PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog

[Changli Gao]

On Fri, May 7, 2010 at 5:51 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> Le vendredi 07 mai 2010 à 07:16 +0200, Eric Dumazet a écrit :
>> Le jeudi 06 mai 2010 à 22:07 -0700, David Miller a écrit :
>>
>> > Looks great, applied, thanks Eric.
>>
>> Thanks, I have a followup to avoid one atomic in enqueue phase too ;)
>>
>
> [PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog
>
> If CONFIG_SMP=y, then we own a queue spinlock, we can avoid the atomic
> test_and_set_bit() from napi_schedule_prep().
>
> We now have same number of atomic ops per netif_rx() calls than with
> pre-RPS kernel.
>
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
> ---
> diff --git a/net/core/dev.c b/net/core/dev.c
> index 32611c8..49fa5a6 100644
> --- a/net/core/dev.c
> +++ b/net/core/dev.c
> @@ -2426,8 +2426,10 @@ enqueue:
>                        return NET_RX_SUCCESS;
>                }
>
> -               /* Schedule NAPI for backlog device */
> -               if (napi_schedule_prep(&sd->backlog)) {
> +               /* Schedule NAPI for backlog device
> +                * We can use non atomic operation since we own the queue lock
> +                */
> +               if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
>                        if (!rps_ipi_queued(sd))
>                                ____napi_schedule(sd, &sd->backlog);
>                }
>

Why not use a wrapper function?

sth. like:

static inline int __napi_schedule_prep(struct napi_struct *n)
{
   return (!__test_and_set_bit(NAPI_STATE_SCHED, &n->state)
}

-- 
Regards，
Changli Gao(xiaosuo@gmail.com)

Re: [PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog

[Eric Dumazet]

Le vendredi 07 mai 2010 à 18:01 +0800, Changli Gao a écrit :
> On Fri, May 7, 2010 at 5:51 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > Le vendredi 07 mai 2010 à 07:16 +0200, Eric Dumazet a écrit :
> >> Le jeudi 06 mai 2010 à 22:07 -0700, David Miller a écrit :
> >>
> >> > Looks great, applied, thanks Eric.
> >>
> >> Thanks, I have a followup to avoid one atomic in enqueue phase too ;)
> >>
> >
> > [PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog
> >
> > If CONFIG_SMP=y, then we own a queue spinlock, we can avoid the atomic
> > test_and_set_bit() from napi_schedule_prep().
> >
> > We now have same number of atomic ops per netif_rx() calls than with
> > pre-RPS kernel.
> >
> > Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
> > ---
> > diff --git a/net/core/dev.c b/net/core/dev.c
> > index 32611c8..49fa5a6 100644
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> > @@ -2426,8 +2426,10 @@ enqueue:
> >                        return NET_RX_SUCCESS;
> >                }
> >
> > -               /* Schedule NAPI for backlog device */
> > -               if (napi_schedule_prep(&sd->backlog)) {
> > +               /* Schedule NAPI for backlog device
> > +                * We can use non atomic operation since we own the queue lock
> > +                */
> > +               if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
> >                        if (!rps_ipi_queued(sd))
> >                                ____napi_schedule(sd, &sd->backlog);
> >                }
> >
> 
> Why not use a wrapper function?
> 
> sth. like:
> 
> static inline int __napi_schedule_prep(struct napi_struct *n)
> {
>    return (!__test_and_set_bit(NAPI_STATE_SCHED, &n->state)
> }
> 


For one user ? 
Not sure it helps code readability.

Right now we all have our minds knowing every bit of this code, but next
year ?

Re: [PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Fri, 07 May 2010 11:51:21 +0200

> Le vendredi 07 mai 2010 à 07:16 +0200, Eric Dumazet a écrit :
>> Le jeudi 06 mai 2010 à 22:07 -0700, David Miller a écrit :
>> 
>> > Looks great, applied, thanks Eric.
>> 
>> Thanks, I have a followup to avoid one atomic in enqueue phase too ;)
>> 
> 
> [PATCH net-next-2.6] rps: avoid one atomic in enqueue_to_backlog
> 
> If CONFIG_SMP=y, then we own a queue spinlock, we can avoid the atomic
> test_and_set_bit() from napi_schedule_prep().
> 
> We now have same number of atomic ops per netif_rx() calls than with
> pre-RPS kernel.
> 
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>

Also applied, thanks Eric.