Hello Maria, Thank you for the answer and for pointing to the ip-iface-inactive branch. I looked at both approaches: the one in your branch and a minimal workaround we prepared on our side. As a result, I have prepared two patches, attached. If you have a chance to review them, I would be grateful for any feedback. 1) 0001 - for ip-iface-inactive (2.19.0 based). While building and testing the branch on our setup, I noticed a couple of things around moving interfaces between the two lists that might be worth a look. I may well be missing some context, so please correct me if so: - an interface that comes back without IF_CHANGE_TOO_MUCH (the same index, e.g. after a device protocol restart) seems to stay in dead_iface_list, so it is not visible to lookups and scans; - if_end_update() moves entries to dead_iface_list while walking iface_list with WALK_LIST(), so the walk seems to stop after the first disappeared interface. In a network namespace with 21 dummy interfaces, after "disable device1" / "enable device1" BIRD showed 20 of 21 interfaces without the patch (the missing one stayed hidden even after "configure"), and 21 of 21 with it. 2) 0002 - an interim workaround for thread-next, in case it is useful before the planned rework. It only unlinks shut-down interfaces nobody refers to (uc == 0, no addresses, no neighbors) and does not free anything, so stale pointers stay valid. Tested on 3.4.0 in a network namespace: - 2000 interfaces with unique names created and deleted: 2001 dead entries in "dump interfaces" without the patch, 1 with it (the pending VRF interface from the config, kept on purpose); - name reuse and a VRF referenced before it exists work as before. I understand both may not fit your plans for the interface code, but if they happen to be of any help, I’d be glad. Best regards, Nikita Koptev пт, 13 нояб. 2026 г. в 19:00, Maria Matejka <maria.matejka@nic.cz>:
Hello!
It's expected to have a completely different data structure for interfaces, together with some larger reworks, but I'm afraid that it's gonna land somewhere around BIRD 3.7 or later.
There may be some workaround inbetween which may help some cases, yet even that is not much expected to get backported. The data structure is almost the original one, expecting that interfaces appearing and disappearing are not a usual thing to happen. A code in progress, with nothing guaranteed, may be in the branch ip-iface-inactive.
I hope this helps. Maria
On November 13, 2026 4:21:25 PM GMT+01:00, Nick_Jeniston < herbby4@gmail.com> wrote:
Hello!
We run BIRD 3.1.3 on hosts where network interfaces (tap, veth, bridges, OVS ports) with unique names are created and deleted all the time. On long-running instances the main I/O loop starts stalling for several seconds ("I/O loop cycle took 5000-6500 ms"). We traced it to removed interfaces that stay in global_iface_list forever: after several months a host has tens of thousands of dead struct iface records (IF_SHUTDOWN, named after long-gone ports), and if_find_by_index / if_update walk all of them. A BIRD restart brings the link scan back from seconds to tens of milliseconds.
This looks like the issue discussed in July 2022: https://bird.network.cz/pipermail/bird-users/2022-July/016228.html
There it was mentioned that reference counting for interface objects would probably land in a future version of BIRD 3. In 3.1.3 if_unlink() still has "TODO: Do some interface object cleanup".
We could not find anything about this in the release announcements since 2022, so we would like to ask:
We could not find any news about this since 2022, so a few questions:
1. Is this issue still on your radar, and is a fix planned? 2. If not, could we contribute by preparing a patch? We would be glad to work on it, following the approach you would prefer. 3. Until then, is there any workaround besides restarting BIRD?
Thank you for developing and maintaining BIRD, it has served us well for years.
Best regards, Nikita Koptev
-- Maria Matejka (she/her) | BIRD Team Leader | CZ.NIC, z.s.p.o.