Re: [PATCH] NETLABEL: Fix an RCU warning

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Thu, Mar 25, 2010 at 12:28:04PM +0100, Eric Dumazet wrote:
> Le jeudi 25 mars 2010 à 11:06 +0000, David Howells a écrit :
> > Fix an RCU warning in the netlabel code due to missing rcu read locking around
> > an rcu_dereference() in netlbl_unlhsh_hash() when called from
> > netlbl_unlhsh_netdev_handler():
> > 
> > ===================================================
> > [ INFO: suspicious rcu_dereference_check() usage. ]
> > ---------------------------------------------------
> > net/netlabel/netlabel_unlabeled.c:246 invoked rcu_dereference_check() without protection!
> > 
> > other info that might help us debug this:
> > 
> > 
> > rcu_scheduler_active = 1, debug_locks = 0
> > 2 locks held by ip/5108:
> >  #0:  (rtnl_mutex){+.+.+.}, at: [<ffffffff812c4a36>] rtnl_lock+0x12/0x14
> >  #1:  (netlbl_unlhsh_lock){+.+...}, at: [<ffffffff8134daa4>] netlbl_unlhsh_netdev_handler+0x1e/0x86
> > 
> > stack backtrace:
> > Pid: 5108, comm: ip Not tainted 2.6.34-rc2-cachefs #114
> > Call Trace:
> >  [<ffffffff8105121f>] lockdep_rcu_dereference+0xaa/0xb2
> >  [<ffffffff8134d781>] netlbl_unlhsh_hash+0x3e/0x50
> >  [<ffffffff8134d7a1>] netlbl_unlhsh_search_iface+0xe/0x84
> >  [<ffffffff8134daaf>] netlbl_unlhsh_netdev_handler+0x29/0x86
> >  [<ffffffff81048362>] notifier_call_chain+0x32/0x5e
> >  [<ffffffff810483fe>] raw_notifier_call_chain+0xf/0x11
> >  [<ffffffff812ba924>] call_netdevice_notifiers+0x16/0x18
> >  [<ffffffff812bac22>] __dev_notify_flags+0x37/0x5b
> >  [<ffffffff812bac8c>] dev_change_flags+0x46/0x52
> >  [<ffffffff812c41af>] do_setlink+0x250/0x3cd
> >  [<ffffffff812c4ee8>] rtnl_newlink+0x2b6/0x49d
> >  [<ffffffff812c4cdd>] ? rtnl_newlink+0xab/0x49d
> >  [<ffffffff812c4c17>] rtnetlink_rcv_msg+0x1b7/0x1d2
> >  [<ffffffff812c4a60>] ? rtnetlink_rcv_msg+0x0/0x1d2
> >  [<ffffffff812cc1dc>] netlink_rcv_skb+0x3e/0x8f
> >  [<ffffffff812c4a59>] rtnetlink_rcv+0x21/0x28
> >  [<ffffffff812cbefe>] netlink_unicast+0x218/0x28f
> >  [<ffffffff812cc76e>] netlink_sendmsg+0x26b/0x27a
> >  [<ffffffff812a9f1d>] sock_sendmsg+0xd4/0xf5
> >  [<ffffffff81096dca>] ? might_fault+0x4e/0x9e
> >  [<ffffffff81096dca>] ? might_fault+0x4e/0x9e
> >  [<ffffffff81096e13>] ? might_fault+0x97/0x9e
> >  [<ffffffff81096dca>] ? might_fault+0x4e/0x9e
> >  [<ffffffff812b4122>] ? verify_iovec+0x59/0x97
> >  [<ffffffff812aa1d9>] sys_sendmsg+0x209/0x273
> >  [<ffffffff810976dc>] ? __do_fault+0x395/0x3cd
> >  [<ffffffff8109928f>] ? handle_mm_fault+0x324/0x69d
> >  [<ffffffff81051e4e>] ? trace_hardirqs_on_caller+0x10c/0x130
> >  [<ffffffff81074972>] ? audit_syscall_entry+0x17d/0x1b0
> >  [<ffffffff81364154>] ? trace_hardirqs_on_thunk+0x3a/0x3f
> >  [<ffffffff81001eeb>] system_call_fastpath+0x16/0x1b
> > 
> > Signed-off-by: David Howells <dhowells@redhat.com>
> > ---
> > 
> >  net/netlabel/netlabel_unlabeled.c |    2 ++
> >  1 files changed, 2 insertions(+), 0 deletions(-)
> > 
> > diff --git a/net/netlabel/netlabel_unlabeled.c b/net/netlabel/netlabel_unlabeled.c
> > index 852d9d7..7ea64e4 100644
> > --- a/net/netlabel/netlabel_unlabeled.c
> > +++ b/net/netlabel/netlabel_unlabeled.c
> > @@ -799,6 +799,7 @@ static int netlbl_unlhsh_netdev_handler(struct notifier_block *this,
> >  
> >  	/* XXX - should this be a check for NETDEV_DOWN or _UNREGISTER? */
> >  	if (event == NETDEV_DOWN) {
> > +		rcu_read_lock();
> >  		spin_lock(&netlbl_unlhsh_lock);
> >  		iface = netlbl_unlhsh_search_iface(dev->ifindex);
> >  		if (iface != NULL && iface->valid) {
> > @@ -807,6 +808,7 @@ static int netlbl_unlhsh_netdev_handler(struct notifier_block *this,
> >  		} else
> >  			iface = NULL;
> >  		spin_unlock(&netlbl_unlhsh_lock);
> > +		rcu_read_unlock();
> >  	}
> >  
> >  	if (iface != NULL)
> > 
> > --
> 
> Sorry this is not the right fix.
> 
> Fix is to change the dereference check to take into account the lock
> owned here.

So we need the rcu_dereference() in netlbl_unlhsh_search_iface()
to become someething like the following?

	bkt_list = &rcu_dereference_check(netlbl_unlhsh,
					  rcu_read_lock_held() ||
					  lockdep_is_held(&netlbl_unlhsh_lock))->tbl[bkt];

Or is this the wrong lock?

							Thanx, Paul