Re: [PATCH net-next v2 1/2] inetpeer: add namespace support for inetpeer

On Tue, Jun 05, 2012 at 02:00:30PM +0200, Eric Dumazet wrote:

On Tue, 2012-06-05 at 13:27 +0200, Steffen Klassert wrote:

quoted

quoted

By the way, I think we have a bug in inetpeer_gc_worker()

Steffen ?

We have no rcu grace period to make sure the following is safe :

if (!atomic_read(&p->refcnt)) {
	list_del(&p->gc_list);
	kmem_cache_free(peer_cachep, p);
}

I think this is ok as it is. inetpeer_invalidate_tree()
unlinks the whole inetpeer tree from the inetpeer base and
adds it to a gc_list. These intetpeer entries are stale,
they can't be looked up again. So noone should increment the
refcount, they just wait until the refcount get zero.

Its not OK, lookups are done under rcu.

Since there is no RCU grace period, the worker free the entries while
another cpus are doing their lookups.

Lookups are done under rcu, yes. But a lookup will not find
these stale entries because the whole interpeer tree is removed
from the inetpeer base before the worker is scheduled. A lookup
would have to cerate a new inetpeer entry in this case.