On Wed, May 25, 2011 at 04:44:29AM +0200, Eric Dumazet wrote:

Hmm, thanks for the report. Are you running x86 or another arch ?

This was on x86.

We probably need some sort of memory barrier.

However, locking this central lock makes the thing too slow, I'll try to
use an atomic_inc_return on p->refcnt instead. (and then lock
unused_peers.lock if we got a 0->1 transition)

Another possibility is to do the list_empty() check twice. Once without
taking the lock and again with the spinlock held.

 -Arun