On Wed, Feb 3, 2016 at 12:08 AM, Julian Anastasov <[email protected]> wrote:
> Aha, I see, it is after NETDEV_UNREGISTER but may be
> the above loop should be changed to two loops so that
> NETDEV_UNREGISTER_BATCH is called exactly after all
> NETDEV_UNREGISTER and before all dev_*_flush and
> ndo_uninit calls to avoid any risks. I mean:
>
> synchronize_net();
>
> First part of loop:
>
> list_for_each_entry(dev, head, unreg_list) {
> /* Shutdown queueing discipline. */
> dev_shutdown(dev);
>
> /* Notify protocols, that we are about to destroy
> this device. They should clean all the things.
> */
> call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
> }
>
> This is the same NETDEV_UNREGISTER_BATCH logic:
>
> + list_for_each_entry(dev, head, unreg_list) {
> + net_add_event_list(&net_head, dev_net(dev));
> + }
> + list_for_each_entry_safe(net, net_tmp, &net_head, event_list) {
> + call_netdevice_notifiers(NETDEV_UNREGISTER_BATCH,
> + net->loopback_dev);
> + net_del_event_list(net);
> + }
>
> Second part of the loop:
>
> list_for_each_entry(dev, head, unreg_list) {
> struct sk_buff *skb = NULL;
>
> if (!dev->rtnl_link_ops ||
> ...
>
> Regards
>
> --
> Julian Anastasov <[email protected]>
You're right. It is probably safer to organize the code the way you said.
Will change that, Thanks!
Salam
