[PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

[PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

[Eric Dumazet]

Now cmpxchg() is available on all arches, we can use it in
build_ehash_secret() and rt_bind_peer() instead of using spinlocks.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
---
 net/ipv4/af_inet.c |    8 +++-----
 net/ipv4/route.c   |   11 +++--------
 2 files changed, 6 insertions(+), 13 deletions(-)

diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
index 6a1100c..f581f77 100644
--- a/net/ipv4/af_inet.c
+++ b/net/ipv4/af_inet.c
@@ -227,18 +227,16 @@ EXPORT_SYMBOL(inet_ehash_secret);
 
 /*
  * inet_ehash_secret must be set exactly once
- * Instead of using a dedicated spinlock, we (ab)use inetsw_lock
  */
 void build_ehash_secret(void)
 {
 	u32 rnd;
+
 	do {
 		get_random_bytes(&rnd, sizeof(rnd));
 	} while (rnd == 0);
-	spin_lock_bh(&inetsw_lock);
-	if (!inet_ehash_secret)
-		inet_ehash_secret = rnd;
-	spin_unlock_bh(&inetsw_lock);
+
+	cmpxchg(&inet_ehash_secret, 0, rnd);
 }
 EXPORT_SYMBOL(build_ehash_secret);
 
diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index 3f56b6e..ae3dba7 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -1268,18 +1268,13 @@ skip_hashing:
 
 void rt_bind_peer(struct rtable *rt, int create)
 {
-	static DEFINE_SPINLOCK(rt_peer_lock);
 	struct inet_peer *peer;
 
 	peer = inet_getpeer(rt->rt_dst, create);
 
-	spin_lock_bh(&rt_peer_lock);
-	if (rt->peer == NULL) {
-		rt->peer = peer;
-		peer = NULL;
-	}
-	spin_unlock_bh(&rt_peer_lock);
-	if (peer)
+	peer = cmpxchg(&rt->peer, NULL, peer);
+
+	if (unlikely(peer))
 		inet_putpeer(peer);
 }
 

Re: [PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

[Changli Gao]

On Thu, Aug 19, 2010 at 11:46 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> Now cmpxchg() is available on all arches, we can use it in
> build_ehash_secret() and rt_bind_peer() instead of using spinlocks.
>
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
> CC: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
> ---
>  net/ipv4/af_inet.c |    8 +++-----
>  net/ipv4/route.c   |   11 +++--------
>  2 files changed, 6 insertions(+), 13 deletions(-)
>
> diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
> index 6a1100c..f581f77 100644
> --- a/net/ipv4/af_inet.c
> +++ b/net/ipv4/af_inet.c
> @@ -227,18 +227,16 @@ EXPORT_SYMBOL(inet_ehash_secret);
>
>  /*
>  * inet_ehash_secret must be set exactly once
> - * Instead of using a dedicated spinlock, we (ab)use inetsw_lock
>  */
>  void build_ehash_secret(void)
>  {
>        u32 rnd;
> +
>        do {
>                get_random_bytes(&rnd, sizeof(rnd));
>        } while (rnd == 0);
> -       spin_lock_bh(&inetsw_lock);
> -       if (!inet_ehash_secret)
> -               inet_ehash_secret = rnd;
> -       spin_unlock_bh(&inetsw_lock);
> +
> +       cmpxchg(&inet_ehash_secret, 0, rnd);
>  }
>  EXPORT_SYMBOL(build_ehash_secret);
>
> diff --git a/net/ipv4/route.c b/net/ipv4/route.c
> index 3f56b6e..ae3dba7 100644
> --- a/net/ipv4/route.c
> +++ b/net/ipv4/route.c
> @@ -1268,18 +1268,13 @@ skip_hashing:
>
>  void rt_bind_peer(struct rtable *rt, int create)
>  {
> -       static DEFINE_SPINLOCK(rt_peer_lock);
>        struct inet_peer *peer;
>
>        peer = inet_getpeer(rt->rt_dst, create);
>
> -       spin_lock_bh(&rt_peer_lock);
> -       if (rt->peer == NULL) {
> -               rt->peer = peer;
> -               peer = NULL;
> -       }
> -       spin_unlock_bh(&rt_peer_lock);
> -       if (peer)
> +       peer = cmpxchg(&rt->peer, NULL, peer);
> +
> +       if (unlikely(peer))

It isn't correct, and should be
     if (unlikely(cmpxchg(&rt->peer, NULL, peer) != NULL))

>                inet_putpeer(peer);
>  }


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

[PATCH net-next-2.6 v2] net: build_ehash_secret() and rt_bind_peer() cleanups

[Eric Dumazet]

Le jeudi 19 août 2010 à 23:58 +0800, Changli Gao a écrit :

> > -       if (peer)
> > +       peer = cmpxchg(&rt->peer, NULL, peer);
> > +
> > +       if (unlikely(peer))
> 
> It isn't correct, and should be
>      if (unlikely(cmpxchg(&rt->peer, NULL, peer) != NULL))
> 
> >                inet_putpeer(peer);
> >  }
> 
> 

Good catch, thanks !

I was trying to take into account peer being NULL...

[PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

Now cmpxchg() is available on all arches, we can use it in
build_ehash_secret() and rt_bind_peer() instead of using spinlocks.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
CC: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
---
 net/ipv4/af_inet.c |    8 +++-----
 net/ipv4/route.c   |    9 +--------
 2 files changed, 4 insertions(+), 13 deletions(-)

diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
index 6a1100c..f581f77 100644
--- a/net/ipv4/af_inet.c
+++ b/net/ipv4/af_inet.c
@@ -227,18 +227,16 @@ EXPORT_SYMBOL(inet_ehash_secret);
 
 /*
  * inet_ehash_secret must be set exactly once
- * Instead of using a dedicated spinlock, we (ab)use inetsw_lock
  */
 void build_ehash_secret(void)
 {
 	u32 rnd;
+
 	do {
 		get_random_bytes(&rnd, sizeof(rnd));
 	} while (rnd == 0);
-	spin_lock_bh(&inetsw_lock);
-	if (!inet_ehash_secret)
-		inet_ehash_secret = rnd;
-	spin_unlock_bh(&inetsw_lock);
+
+	cmpxchg(&inet_ehash_secret, 0, rnd);
 }
 EXPORT_SYMBOL(build_ehash_secret);
 
diff --git a/net/ipv4/route.c b/net/ipv4/route.c
index 3f56b6e..85a67c9 100644
--- a/net/ipv4/route.c
+++ b/net/ipv4/route.c
@@ -1268,18 +1268,11 @@ skip_hashing:
 
 void rt_bind_peer(struct rtable *rt, int create)
 {
-	static DEFINE_SPINLOCK(rt_peer_lock);
 	struct inet_peer *peer;
 
 	peer = inet_getpeer(rt->rt_dst, create);
 
-	spin_lock_bh(&rt_peer_lock);
-	if (rt->peer == NULL) {
-		rt->peer = peer;
-		peer = NULL;
-	}
-	spin_unlock_bh(&rt_peer_lock);
-	if (peer)
+	if (peer && cmpxchg(&rt->peer, NULL, peer) != NULL)
 		inet_putpeer(peer);
 }
 

Re: [PATCH net-next-2.6 v2] net: build_ehash_secret() and rt_bind_peer() cleanups

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Thu, 19 Aug 2010 18:10:45 +0200

> [PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups
> 
> Now cmpxchg() is available on all arches, we can use it in
> build_ehash_secret() and rt_bind_peer() instead of using spinlocks.
> 
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>

Applied, thanks Eric.

Re: [PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

[Changli Gao]

On Thu, Aug 19, 2010 at 11:58 PM, Changli Gao <xiaosuo@gmail.com> wrote:
> On Thu, Aug 19, 2010 at 11:46 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
>> Now cmpxchg() is available on all arches, we can use it in
>> build_ehash_secret() and rt_bind_peer() instead of using spinlocks.
>>
>> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
>> CC: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
>> ---
>>  net/ipv4/af_inet.c |    8 +++-----
>>  net/ipv4/route.c   |   11 +++--------
>>  2 files changed, 6 insertions(+), 13 deletions(-)
>>
>> diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
>> index 6a1100c..f581f77 100644
>> --- a/net/ipv4/af_inet.c
>> +++ b/net/ipv4/af_inet.c
>> @@ -227,18 +227,16 @@ EXPORT_SYMBOL(inet_ehash_secret);
>>
>>  /*
>>  * inet_ehash_secret must be set exactly once
>> - * Instead of using a dedicated spinlock, we (ab)use inetsw_lock
>>  */
>>  void build_ehash_secret(void)
>>  {
>>        u32 rnd;
>> +
>>        do {
>>                get_random_bytes(&rnd, sizeof(rnd));
>>        } while (rnd == 0);
>> -       spin_lock_bh(&inetsw_lock);
>> -       if (!inet_ehash_secret)
>> -               inet_ehash_secret = rnd;
>> -       spin_unlock_bh(&inetsw_lock);
>> +
>> +       cmpxchg(&inet_ehash_secret, 0, rnd);
>>  }
>>  EXPORT_SYMBOL(build_ehash_secret);
>>
>> diff --git a/net/ipv4/route.c b/net/ipv4/route.c
>> index 3f56b6e..ae3dba7 100644
>> --- a/net/ipv4/route.c
>> +++ b/net/ipv4/route.c
>> @@ -1268,18 +1268,13 @@ skip_hashing:
>>
>>  void rt_bind_peer(struct rtable *rt, int create)
>>  {
>> -       static DEFINE_SPINLOCK(rt_peer_lock);
>>        struct inet_peer *peer;
>>
>>        peer = inet_getpeer(rt->rt_dst, create);
>>
>> -       spin_lock_bh(&rt_peer_lock);
>> -       if (rt->peer == NULL) {
>> -               rt->peer = peer;
>> -               peer = NULL;
>> -       }
>> -       spin_unlock_bh(&rt_peer_lock);
>> -       if (peer)
>> +       peer = cmpxchg(&rt->peer, NULL, peer);
>> +
>> +       if (unlikely(peer))
>
> It isn't correct, and should be
>     if (unlikely(cmpxchg(&rt->peer, NULL, peer) != NULL))
>

we should also test peer != NULL.


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

Re: [PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

[Mathieu Desnoyers]

* Changli Gao (xiaosuo@gmail.com) wrote:
> On Thu, Aug 19, 2010 at 11:46 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> > Now cmpxchg() is available on all arches, we can use it in
> > build_ehash_secret() and rt_bind_peer() instead of using spinlocks.
> >
> > Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
> > CC: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
> > ---
[...]
> >  void rt_bind_peer(struct rtable *rt, int create)
> >  {
> > -       static DEFINE_SPINLOCK(rt_peer_lock);
> >        struct inet_peer *peer;
> >
> >        peer = inet_getpeer(rt->rt_dst, create);
> >
> > -       spin_lock_bh(&rt_peer_lock);
> > -       if (rt->peer == NULL) {
> > -               rt->peer = peer;
> > -               peer = NULL;
> > -       }
> > -       spin_unlock_bh(&rt_peer_lock);
> > -       if (peer)
> > +       peer = cmpxchg(&rt->peer, NULL, peer);
> > +
> > +       if (unlikely(peer))
> 
> It isn't correct, and should be
>      if (unlikely(cmpxchg(&rt->peer, NULL, peer) != NULL))
> 
> >                inet_putpeer(peer);
> >  }

What do you mean by "It isn't correct" ? Eric code and yours perform the
exact same thing. So I assume this is just a code style issue here.

Thanks,

Mathieu


-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com

Re: [PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

[Mathieu Desnoyers]

* Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca) wrote:
> * Changli Gao (xiaosuo@gmail.com) wrote:
> > On Thu, Aug 19, 2010 at 11:46 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
[...]
> > > +       peer = cmpxchg(&rt->peer, NULL, peer);
> > > +
> > > +       if (unlikely(peer))
> > 
> > It isn't correct, and should be
> >      if (unlikely(cmpxchg(&rt->peer, NULL, peer) != NULL))
> > 
> > >                inet_putpeer(peer);
> > >  }
> 
> What do you mean by "It isn't correct" ? Eric code and yours perform the
> exact same thing. So I assume this is just a code style issue here.

Nevermind my comment, it's not accurate.

Thanks,

Mathieu


-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com

Re: [PATCH net-next-2.6] net: build_ehash_secret() and rt_bind_peer() cleanups

[Mathieu Desnoyers]

* Eric Dumazet (eric.dumazet@gmail.com) wrote:
> Now cmpxchg() is available on all arches, we can use it in
> build_ehash_secret() and rt_bind_peer() instead of using spinlocks.
> 
> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
> CC: Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
> ---
>  net/ipv4/af_inet.c |    8 +++-----
>  net/ipv4/route.c   |   11 +++--------
>  2 files changed, 6 insertions(+), 13 deletions(-)
> 
> diff --git a/net/ipv4/af_inet.c b/net/ipv4/af_inet.c
> index 6a1100c..f581f77 100644
> --- a/net/ipv4/af_inet.c
> +++ b/net/ipv4/af_inet.c
> @@ -227,18 +227,16 @@ EXPORT_SYMBOL(inet_ehash_secret);
>  
>  /*
>   * inet_ehash_secret must be set exactly once
> - * Instead of using a dedicated spinlock, we (ab)use inetsw_lock
>   */
>  void build_ehash_secret(void)
>  {
>  	u32 rnd;
> +
>  	do {
>  		get_random_bytes(&rnd, sizeof(rnd));
>  	} while (rnd == 0);
> -	spin_lock_bh(&inetsw_lock);
> -	if (!inet_ehash_secret)
> -		inet_ehash_secret = rnd;
> -	spin_unlock_bh(&inetsw_lock);
> +
> +	cmpxchg(&inet_ehash_secret, 0, rnd);

I'd be more comfortable if you add a comment saying why you can get away
with not testing the cmpxchg failure case. (it's because the failure is
caused by a concurrent CPU setting the ehash secret to a random value,
which actually does the job for us, so no need to retry).

>  }
>  EXPORT_SYMBOL(build_ehash_secret);
>  
> diff --git a/net/ipv4/route.c b/net/ipv4/route.c
> index 3f56b6e..ae3dba7 100644
> --- a/net/ipv4/route.c
> +++ b/net/ipv4/route.c
> @@ -1268,18 +1268,13 @@ skip_hashing:
>  
>  void rt_bind_peer(struct rtable *rt, int create)
>  {
> -	static DEFINE_SPINLOCK(rt_peer_lock);
>  	struct inet_peer *peer;
>  
>  	peer = inet_getpeer(rt->rt_dst, create);

Hrm, peer can be NULL on OOM condition.

>  
> -	spin_lock_bh(&rt_peer_lock);
> -	if (rt->peer == NULL) {
> -		rt->peer = peer;
> -		peer = NULL;
> -	}
> -	spin_unlock_bh(&rt_peer_lock);
> -	if (peer)
> +	peer = cmpxchg(&rt->peer, NULL, peer);
> +
> +	if (unlikely(peer))
>  		inet_putpeer(peer);

And you don't want to put the peer returned by cmpxchg, you want to put
the peer you just allocated.

I'd go for:

  if (!peer)
    return;
  if (unlikely(cmpxchg(&rt->peer, NULL, peer) != NULL))
    inet_putpeer(peer);

Thanks,

Mathieu


>  }
>  
> 
> 

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com