diff mbox

[net-next] ipv6: Avoid rt6_probe() taking writer lock in the fast path

Message ID 1437522680-3515872-1-git-send-email-kafai@fb.com
State Changes Requested, archived
Delegated to: David Miller
Headers show

Commit Message

Martin KaFai Lau July 21, 2015, 11:51 p.m. UTC
The patch checks neigh->nud_state before acquiring the writer lock.
Note that rt6_probe() is only used in CONFIG_IPV6_ROUTER_PREF.

I also take this chance to re-arrange the code.

40 udpflood processes and a /64 gateway route are used.
The gateway has NUD_PERMANENT.  Each of them is run for 30s.
At the end, the total number of finished sendto():

Before    After
55M	  95M

Signed-off-by: Martin KaFai Lau <kafai@fb.com>
Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
---
 net/ipv6/route.c | 41 ++++++++++++++++++++---------------------
 1 file changed, 20 insertions(+), 21 deletions(-)

Comments

Hideaki Yoshifuji July 22, 2015, 2:10 a.m. UTC | #1
Hi,

Martin KaFai Lau wrote:
> The patch checks neigh->nud_state before acquiring the writer lock.
> Note that rt6_probe() is only used in CONFIG_IPV6_ROUTER_PREF.

You have to take "some" lock when accessing neigh->nud_state
theoretically.

> 
> I also take this chance to re-arrange the code.

No, please do not mix multiple changes.

> 
> 40 udpflood processes and a /64 gateway route are used.
> The gateway has NUD_PERMANENT.  Each of them is run for 30s.
> At the end, the total number of finished sendto():
> 
> Before    After
> 55M	  95M
> 
> Signed-off-by: Martin KaFai Lau <kafai@fb.com>
> Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
> ---
>  net/ipv6/route.c | 41 ++++++++++++++++++++---------------------
>  1 file changed, 20 insertions(+), 21 deletions(-)
> 
> diff --git a/net/ipv6/route.c b/net/ipv6/route.c
> index 6090969..a6c6b5a 100644
> --- a/net/ipv6/route.c
> +++ b/net/ipv6/route.c
> @@ -544,6 +544,7 @@ static void rt6_probe_deferred(struct work_struct *w)
>  
>  static void rt6_probe(struct rt6_info *rt)
>  {
> +	struct __rt6_probe_work *work;
>  	struct neighbour *neigh;
>  	/*
>  	 * Okay, this does not seem to be appropriate
> @@ -558,34 +559,32 @@ static void rt6_probe(struct rt6_info *rt)
>  	rcu_read_lock_bh();
>  	neigh = __ipv6_neigh_lookup_noref(rt->dst.dev, &rt->rt6i_gateway);
>  	if (neigh) {
> -		write_lock(&neigh->lock);
>  		if (neigh->nud_state & NUD_VALID)
>  			goto out;
> -	}
> -
> -	if (!neigh ||
> -	    time_after(jiffies, neigh->updated + rt->rt6i_idev->cnf.rtr_probe_interval)) {
> -		struct __rt6_probe_work *work;
>  
> +		work = NULL;
> +		write_lock(&neigh->lock);
> +		if (!(neigh->nud_state & NUD_VALID) &&
> +		    time_after(jiffies, neigh->updated + rt->rt6i_idev->cnf.rtr_probe_interval)) {
> +			work = kmalloc(sizeof(*work), GFP_ATOMIC);
> +			if (work) {
> +				__neigh_set_probe_once(neigh);
> +			}
> +		}
> +		write_unlock(&neigh->lock);
> +	} else {
>  		work = kmalloc(sizeof(*work), GFP_ATOMIC);
> +	}
>  
> -		if (neigh && work)
> -			__neigh_set_probe_once(neigh);
> -
> -		if (neigh)
> -			write_unlock(&neigh->lock);
> +	if (work) {
> +		INIT_WORK(&work->work, rt6_probe_deferred);
> +		work->target = rt->rt6i_gateway;
> +		dev_hold(rt->dst.dev);
> +		work->dev = rt->dst.dev;
> +		schedule_work(&work->work);
> +	}
>  
> -		if (work) {
> -			INIT_WORK(&work->work, rt6_probe_deferred);
> -			work->target = rt->rt6i_gateway;
> -			dev_hold(rt->dst.dev);
> -			work->dev = rt->dst.dev;
> -			schedule_work(&work->work);
> -		}
> -	} else {
>  out:
> -		write_unlock(&neigh->lock);
> -	}
>  	rcu_read_unlock_bh();
>  }
>  #else
>
Julian Anastasov July 22, 2015, 4:19 a.m. UTC | #2
Hello,

On Tue, 21 Jul 2015, Martin KaFai Lau wrote:

> The patch checks neigh->nud_state before acquiring the writer lock.
> Note that rt6_probe() is only used in CONFIG_IPV6_ROUTER_PREF.

	Locking usage is absolutely correct.

> +		if (!(neigh->nud_state & NUD_VALID) &&
> +		    time_after(jiffies, neigh->updated + rt->rt6i_idev->cnf.rtr_probe_interval)) {

	but this line is too long...

> +			work = kmalloc(sizeof(*work), GFP_ATOMIC);
> +			if (work) {
> +				__neigh_set_probe_once(neigh);
> +			}

	scripts/checkpatch.pl --strict /tmp/file.patch

Regards

--
Julian Anastasov <ja@ssi.bg>
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Martin KaFai Lau July 22, 2015, 6:43 p.m. UTC | #3
On Wed, Jul 22, 2015 at 11:10:59AM +0900, YOSHIFUJI Hideaki wrote:
> You have to take "some" lock when accessing neigh->nud_state
> theoretically.
I don't think read_lock can buy us a lot of extra protection either.
If it has missed the train, the next ip6_pol_route() call will
trigger rt6_probe().  There are other places checking nud_state without
lock also.

Thanks,
--Martin
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index 6090969..a6c6b5a 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -544,6 +544,7 @@  static void rt6_probe_deferred(struct work_struct *w)
 
 static void rt6_probe(struct rt6_info *rt)
 {
+	struct __rt6_probe_work *work;
 	struct neighbour *neigh;
 	/*
 	 * Okay, this does not seem to be appropriate
@@ -558,34 +559,32 @@  static void rt6_probe(struct rt6_info *rt)
 	rcu_read_lock_bh();
 	neigh = __ipv6_neigh_lookup_noref(rt->dst.dev, &rt->rt6i_gateway);
 	if (neigh) {
-		write_lock(&neigh->lock);
 		if (neigh->nud_state & NUD_VALID)
 			goto out;
-	}
-
-	if (!neigh ||
-	    time_after(jiffies, neigh->updated + rt->rt6i_idev->cnf.rtr_probe_interval)) {
-		struct __rt6_probe_work *work;
 
+		work = NULL;
+		write_lock(&neigh->lock);
+		if (!(neigh->nud_state & NUD_VALID) &&
+		    time_after(jiffies, neigh->updated + rt->rt6i_idev->cnf.rtr_probe_interval)) {
+			work = kmalloc(sizeof(*work), GFP_ATOMIC);
+			if (work) {
+				__neigh_set_probe_once(neigh);
+			}
+		}
+		write_unlock(&neigh->lock);
+	} else {
 		work = kmalloc(sizeof(*work), GFP_ATOMIC);
+	}
 
-		if (neigh && work)
-			__neigh_set_probe_once(neigh);
-
-		if (neigh)
-			write_unlock(&neigh->lock);
+	if (work) {
+		INIT_WORK(&work->work, rt6_probe_deferred);
+		work->target = rt->rt6i_gateway;
+		dev_hold(rt->dst.dev);
+		work->dev = rt->dst.dev;
+		schedule_work(&work->work);
+	}
 
-		if (work) {
-			INIT_WORK(&work->work, rt6_probe_deferred);
-			work->target = rt->rt6i_gateway;
-			dev_hold(rt->dst.dev);
-			work->dev = rt->dst.dev;
-			schedule_work(&work->work);
-		}
-	} else {
 out:
-		write_unlock(&neigh->lock);
-	}
 	rcu_read_unlock_bh();
 }
 #else