On Wed, Sep 05, 2018 at 04:07:48PM -0600, Parav Pandit wrote: > > Is taking it before calling rdma_find_ndev_for_src_ip_rcu() the right thing to do? > Likely because rcu_read_lock can be nested as per [1]. > If we do rcu_read_lock() inside rdma_find_ndev_for_src_ip_rcu(), we > need to dev_hold and dev_put() and handle code differently between > ipv4 and v6. ?? We already have dev_hold in both and we already dev_put, if we move the read_lock it would be just to put it in the ipv6 leg of that switch.. > However, dev_hold() and dev_put() doesn't guarantee that netdev > fields won't change such as ifindex. It just ensures that netdev > exist. That explanation would make more sense if we didn't immediately drop the rcu after copying the if_index into a struct. Nothing can safely read that structure member, so why did we do it? I thought we got rid of all the ifindex stuff? Is this just a weirdness in-between patches in the series? I haven't got to look at these patches yet.. There is definately something wonky here.. Jason
