I'm working with Rodrigo Vaz and we've found some details about our occurrence of this issue using systemtap.
rt_cache_route places a dst_entry struct into a fib_nh struct as the nh_rth_input. Occasionally the reference counter on that dst is not decremented by the time free_fib_info_rcu is called on the fib during container teardown. In that case free_fib_info_rcu doesn't call dst_destroy and dev_put is not called on the lo interface of the container. The only situation where we've seen this is when A) the fib_nh->nh_dev points to the eth0 interface of the container, B) the dst is part of an rtable struct where rt-> rt_is_input==1, and C) the dst points to the lo interface of the container. The dst is cached in the fib only once and never replaced and then thousands of dst_hold/dst_release calls are made on the dst for connections and socket buffers. We have only seen this so far on containers making lots of outbound connections. It doesn't appear to depend on the lifetime of the container, some are only alive for 30min and others are alive for 24hrs. The issue occurs when you try to destroy the container because that is when the fib is freed. We don't know when or where the dst ref_cnt becomes incorrect. We don't know how to reproduce the issue. -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/1403152 Title: unregister_netdevice: waiting for lo to become free. Usage count To manage notifications about this bug go to: https://bugs.launchpad.net/linux/+bug/1403152/+subscriptions -- ubuntu-bugs mailing list [email protected] https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs
