Am Freitag, den 31.12.2010, 13:00 +0100 schrieb Eric Dumazet:

Le vendredi 31 décembre 2010 à 12:25 +0100, Eric Dumazet a écrit :

quoted

Le vendredi 31 décembre 2010 à 10:29 +0100, stefani@seibold.net a
écrit :

quoted

+		if (!list_empty(&usk->destlist)) {
+			state->sk = (struct sock *)usk;
+			state->dest = list_first_entry(&usk->destlist,
+					struct udpcp_dest, list);
+			sock_hold(state->sk);
+
+			if (atomic_read(&state->sk->sk_refcnt) != 1) {
+				spin_unlock_irqrestore(&spinlock, flags);
+				return state;
+			}
+			atomic_dec(&state->sk->sk_refcnt);
+		}
+

I am trying to understand what you are doing here.

It seems racy to me.

Apparently, what you want is to take a reference only if actual
sk_refcnt is not zero.

I suggest using atomic_inc_notzero(&state->sk->sk_refcnt) to avoid the
race in atomic_dec().

Before you ask why its racy, this is because UDP sockets are RCU
protected, and RCU lookups depend on sk_refcnt being zero or not.

Doing an sk_refcnt increment/decrement opens a race window for the
concurrent lookups.

I still revamped the whole /proc/net/udpcp thing and hope it is now race
free.