On 16.03.2018 16:00, Sowmini Varadhan wrote: > On (03/16/18 15:38), Kirill Tkhai wrote: >> >> 467fa15356acf by Sowmini Varadhan added NETDEV_UNREGISTER_FINAL dependence >> with the commentary: >> >> /* rds-tcp registers as a pernet subys, so the ->exit will only >> * get invoked after network acitivity has quiesced. We need to >> * clean up all sockets to quiesce network activity, and use >> * the unregistration of the per-net loopback device as a trigger >> * to start that cleanup. >> */ >> >> It seems all the protocols pernet subsystems meet this situation, but they >> solve it in generic way. What does rds so specific have? > > The difference with rds is this: most consumers of netns associate > a net_device with the netns, and cleanup of the netns state > happens as part of the net_device teardown without the constraint > above. rds-tcp does has a netns tied to listening socket, not > to a specific network interface (net_device) so it registers > as a pernet-subsys. But this means that cleanup has to be > cone carefully (see comments in net_namespace.h before > register_pernet_subsys)
This is not a problem, and rds-tcp is not the only pernet_subsys registering a socket. It's OK to close it from .exit method. There are many examples, let me point you to icmp_sk_ops as one of them. But it's not the only. > For rds-tcp, we need to be able to do the right thing in both of these > cases > 1. modprobe -r rds-tcp (cleanup of rds-tcp state should happen in > every namespace, including init_net) > 2. netns delete (rds_tcp.ko should remain loaded for other namespaces) The same as above, every pernet_subsys does this. It's not a problem. exit and exit_batch methods are called in both of the cases. Please, see __unregister_pernet_operations()->ops_exit_list for the details. >> This commit makes event handler to iterate rds_tcp_conn_list and >> kill them. If we change the stage to NETDEV_UNREGISTER, what will change? > > The above two cases need to work correctly. Yeah, but let's find another way to have the same. >> Can unregistered loopback device on dead net add new items to >> rds_tcp_conn_list? >> How it's possible? > > I dont understand the question- no unregistered loopback devices > cannot add items. If we replace NETDEV_UNREGISTER_FINAL with NETDEV_UNREGISTER, the only change which happens is we call rds_tcp_kill_sock() earlier. So, it may be a reason of problems only if someone changes the list during the time between NETDEV_UNREGISTER and NETDEV_UNREGISTER_FINAL are called for loopback. But since this time noone related to this net can extend the list, there is no a problem to do that. > fwiw, I had asked questions about this (netns per net_device > vs netns for module) on the netdev list a few years ago, I can > try to hunt down that thread for you later (nobody replied to > it, but maybe it will help answer your questions). After your words it looks like we may simply do all the things in exit method. Kirill
