diff mbox

[RFC] packet: change call of synchronize_net to call_rcu

Message ID 1346768215-5194-1-git-send-email-icurt@ixiacom.com
State RFC, archived
Delegated to: David Miller
Headers show

Commit Message

Iulius Curt Sept. 4, 2012, 2:16 p.m. UTC 
synchronize_net is called every time we close a PF_PACKET socket which is
causing performance loss when doing this on many sockets.

Signed-off-by: Sorin Dumitru <sdumitru@ixiacom.com>
Signed-off-by: Iulius Curt <icurt@ixiacom.com>
---
Statistics using test program [1]

Sockets count  |  Not patched  |  Patched

Comments

David Miller Sept. 4, 2012, 4:38 p.m. UTC | #1 
From: Daniel Borkmann <danborkmann@iogearbox.net>
Date: Tue, 4 Sep 2012 16:53:06 +0200

> On Tue, Sep 4, 2012 at 4:16 PM, Iulius Curt <iulius.curt@gmail.com> wrote:
>> synchronize_net is called every time we close a PF_PACKET socket which is
>> causing performance loss when doing this on many sockets.
> 
> Do you have any particular use case in mind?

I'm curious about this too, it seems rediculous to optimize for this
and that only seriously mis-designed userspace would do something like
this.

--
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/packet/af_packet.c b/net/packet/af_packet.c
index 5dafe84..7b60135 100644
--- a/net/packet/af_packet.c
+++ b/net/packet/af_packet.c
@@ -2299,6 +2299,32 @@  static int packet_sendmsg(struct kiocb *iocb, struct socket *sock,
 		return packet_snd(sock, msg, len);
 }
 
+void packet_release_finish(struct sock *sk, int null)
+{
+	struct socket *sock = sk->sk_socket;
+
+	sock_orphan(sk);
+	/* Modifing this after the socket has been released
+	 * might lead to memory corruption so we don't do it */
+	if (null)
+		sock->sk = NULL;
+
+	/* Purge queues */
+
+	skb_queue_purge(&sk->sk_receive_queue);
+	sk_refcnt_debug_release(sk);
+
+	sock_put(sk);
+}
+
+void packet_release_rcu(struct rcu_head *rcu)
+{
+	struct packet_sock *po = container_of(rcu, struct packet_sock, rcu_head);
+	struct sock *sk = &po->sk;
+
+	packet_release_finish(sk, 0);
+}
+
 /*
  *	Close a PACKET socket. This is fairly simple. We immediately go
  *	to 'closed' state and remove our protocol entry in the device list.
@@ -2310,6 +2336,7 @@  static int packet_release(struct socket *sock)
 	struct packet_sock *po;
 	struct net *net;
 	union tpacket_req_u req_u;
+	int postpone = 0;
 
 	if (!sk)
 		return 0;
@@ -2326,7 +2353,10 @@  static int packet_release(struct socket *sock)
 	preempt_enable();
 
 	spin_lock(&po->bind_lock);
-	unregister_prot_hook(sk, false);
+	if (po->running) {
+		postpone = 1;
+		__unregister_prot_hook(sk, false);
+	}
 	if (po->prot_hook.dev) {
 		dev_put(po->prot_hook.dev);
 		po->prot_hook.dev = NULL;
@@ -2345,19 +2375,15 @@  static int packet_release(struct socket *sock)
 
 	fanout_release(sk);
 
-	synchronize_net();
 	/*
 	 *	Now the socket is dead. No more input will appear.
 	 */
-	sock_orphan(sk);
-	sock->sk = NULL;
-
-	/* Purge queues */
-
-	skb_queue_purge(&sk->sk_receive_queue);
-	sk_refcnt_debug_release(sk);
+	if (postpone) {
+		call_rcu(&po->rcu_head, packet_release_rcu);
+		return 0;
+	}
 
-	sock_put(sk);
+	packet_release_finish(sk, 1);
 	return 0;
 }
 
diff --git a/net/packet/internal.h b/net/packet/internal.h
index 44945f6..5778c12 100644
--- a/net/packet/internal.h
+++ b/net/packet/internal.h
@@ -104,6 +104,7 @@  struct packet_sock {
 	int			ifindex;	/* bound device		*/
 	__be16			num;
 	struct packet_mclist	*mclist;
+	struct rcu_head		rcu_head;
 	atomic_t		mapped;
 	enum tpacket_versions	tp_version;
 	unsigned int		tp_hdrlen;