[PATCH net-next] ipv6: avoid taking locks at socket dismantle

[PATCH net-next] ipv6: avoid taking locks at socket dismantle

[Eric Dumazet]

From: Eric Dumazet <edumazet@google.com>

ipv6_sock_mc_close() is called for ipv6 sockets at close time, and most
of them don't use multicast.

Add a test to avoid contention on a shared spinlock.

Same heuristic applies for ipv6_sock_ac_close(), to avoid contention
on a shared rwlock.

Signed-off-by: Eric Dumazet <edumazet@google.com>
---
 net/ipv6/anycast.c |    3 +++
 net/ipv6/mcast.c   |    3 +++
 2 files changed, 6 insertions(+)

diff --git a/net/ipv6/anycast.c b/net/ipv6/anycast.c
index 2f4f584..757a810 100644
--- a/net/ipv6/anycast.c
+++ b/net/ipv6/anycast.c
@@ -189,6 +189,9 @@ void ipv6_sock_ac_close(struct sock *sk)
 	struct net *net = sock_net(sk);
 	int	prev_index;
 
+	if (!np->ipv6_ac_list)
+		return;
+
 	write_lock_bh(&ipv6_sk_ac_lock);
 	pac = np->ipv6_ac_list;
 	np->ipv6_ac_list = NULL;
diff --git a/net/ipv6/mcast.c b/net/ipv6/mcast.c
index b19ed51..28dfa5f 100644
--- a/net/ipv6/mcast.c
+++ b/net/ipv6/mcast.c
@@ -284,6 +284,9 @@ void ipv6_sock_mc_close(struct sock *sk)
 	struct ipv6_mc_socklist *mc_lst;
 	struct net *net = sock_net(sk);
 
+	if (!rcu_access_pointer(np->ipv6_mc_list))
+		return;
+
 	spin_lock(&ipv6_sk_mc_lock);
 	while ((mc_lst = rcu_dereference_protected(np->ipv6_mc_list,
 				lockdep_is_held(&ipv6_sk_mc_lock))) != NULL) {

Re: [PATCH net-next] ipv6: avoid taking locks at socket dismantle

[David Stevens]

> From: Eric Dumazet <edumazet@google.com>
> 
> ipv6_sock_mc_close() is called for ipv6 sockets at close time, and most
> of them don't use multicast.
> 
> Add a test to avoid contention on a shared spinlock.
> 
> Same heuristic applies for ipv6_sock_ac_close(), to avoid contention
> on a shared rwlock.

        What prevents a different thread from racing with the
tests for NULL on these?

                                                +-DLS

Re: [PATCH net-next] ipv6: avoid taking locks at socket dismantle

[David Miller]

From: David Stevens <dlstevens@us.ibm.com>
Date: Wed, 5 Dec 2012 16:07:16 -0500

>> From: Eric Dumazet <edumazet@google.com>
>> 
>> ipv6_sock_mc_close() is called for ipv6 sockets at close time, and most
>> of them don't use multicast.
>> 
>> Add a test to avoid contention on a shared spinlock.
>> 
>> Same heuristic applies for ipv6_sock_ac_close(), to avoid contention
>> on a shared rwlock.
> 
>         What prevents a different thread from racing with the
> tests for NULL on these?

The socket is being torn apart, which means that operations on it's
FD are no longer possible, which means that no other thread can add
entries to these lists.

Re: [PATCH net-next] ipv6: avoid taking locks at socket dismantle

[David Miller]

From: Eric Dumazet <eric.dumazet@gmail.com>
Date: Wed, 05 Dec 2012 11:18:10 -0800

> From: Eric Dumazet <edumazet@google.com>
> 
> ipv6_sock_mc_close() is called for ipv6 sockets at close time, and most
> of them don't use multicast.
> 
> Add a test to avoid contention on a shared spinlock.
> 
> Same heuristic applies for ipv6_sock_ac_close(), to avoid contention
> on a shared rwlock.
> 
> Signed-off-by: Eric Dumazet <edumazet@google.com>

Applied.