On 08/06/18 16:11, Petr Machata wrote:
> Callers of br_fdb_find() need to hold the hash lock, which
> br_fdb_find_port() doesn't do. However, since br_fdb_find_port() is not
> doing any actual FDB manipulation, the hash lock is not really needed at
> all. So convert to br_fdb_find_rcu(), surrounded by rcu_read_lock() /
> _unlock() pair.
>
> The device pointer copied from inside the FDB entry is then kept alive
> by the RTNL lock, which br_fdb_find_port() asserts.
>
> Fixes: 4d4fd36126d6 ("net: bridge: Publish bridge accessor functions")
> Signed-off-by: Petr Machata <petrm@xxxxxxxxxxxx>
> ---
>
> Notes:
> Changes from v1 to v2:
>
> - Instead of taking hash lock, take RCU lock and call br_fdb_find_rcu().
>
> net/bridge/br_fdb.c | 4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> diff --git a/net/bridge/br_fdb.c b/net/bridge/br_fdb.c
> index b19e310..502f663 100644
> --- a/net/bridge/br_fdb.c
> +++ b/net/bridge/br_fdb.c
> @@ -135,9 +135,11 @@ struct net_device *br_fdb_find_port(const struct net_device *br_dev,
> return NULL;
>
> br = netdev_priv(br_dev);
> - f = br_fdb_find(br, addr, vid);
> + rcu_read_lock();
> + f = br_fdb_find_rcu(br, addr, vid);
> if (f && f->dst)
> dev = f->dst->dev;
> + rcu_read_unlock();
>
> return dev;
> }
>
Important note: the only reason this will not dereference a NULL pointer
when getting f->dst is because RTNL is held in all of its current
callers. I missed the comments on the previous version, but using RCU
here is dangerous if someone decides to use this without rtnl they will
get a false sense of security, that is why I acked the previous version.
I'd suggest to use READ_ONCE() for f->dst to avoid reading it again.
The way to reach a null dst is with fdbs pointing to the bridge which
currently can only be installed via user-space with RTNL held.
