On 01.04.2016 03:45, Eric Dumazet wrote:

On Thu, 2016-03-31 at 18:39 -0700, Eric Dumazet wrote:

quoted

On Fri, 2016-04-01 at 03:36 +0200, Hannes Frederic Sowa wrote:

quoted

On Fri, Apr 1, 2016, at 03:19, Eric Dumazet wrote:

quoted

Thanks.

As you can see, release_sock() messes badly lockdep (once your other
patches are in )

Once we properly fix release_sock() and/or __release_sock(), all these
false positives disappear.

This was a loopback connection. I need to study release_sock and
__release_sock more as I cannot currently see an issue with the lockdep
handling.

Okay, please try :

diff --git a/net/core/sock.c b/net/core/sock.c
index b67b9aedb230..570dcd91d64e 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c

@@ -2429,10 +2429,6 @@ EXPORT_SYMBOL(lock_sock_nested);

  void release_sock(struct sock *sk)
  {
-	/*
-	 * The sk_lock has mutex_unlock() semantics:
-	 */
-	mutex_release(&sk->sk_lock.dep_map, 1, _RET_IP_);

  	spin_lock_bh(&sk->sk_lock.slock);
  	if (sk->sk_backlog.tail)

@@ -2445,6 +2441,10 @@ void release_sock(struct sock *sk)
  		sk->sk_prot->release_cb(sk);

  	sock_release_ownership(sk);
+	/*
+	 * The sk_lock has mutex_unlock() semantics:
+	 */
+	mutex_release(&sk->sk_lock.dep_map, 1, _RET_IP_);
  	if (waitqueue_active(&sk->sk_lock.wq))
  		wake_up(&sk->sk_lock.wq);
  	spin_unlock_bh(&sk->sk_lock.slock);

Also take a look at commit c3f9b01849ef3bc69024990092b9f42e20df7797

We might need to include the mutex_release() in sock_release_ownership()

I thought so first, as well. But given the double check for the 
spin_lock and the "mutex" we end up with the same result for the 
lockdep_sock_is_held check.

Do you see other consequences?

Thanks,
Hannes