APPLIED[E]/cmt: [SRU][B/D] Ensure /proc/sys/net/bridge folders (dis)appear appropriately

Seth Forshee seth.forshee at
Tue Jul 30 16:01:08 UTC 2019

+Cc Christian.

On Thu, Jul 25, 2019 at 05:20:46PM -0700, Connor Kuehl wrote:
> Note: Bionic required two additional patches in order for these to apply cleanly, one
> of which required minor backporting to use the updated wrappers/symbols.
> BugLink:
> Justification taken from the link above ^
> SRU Justification
> Impact: Currently, the /proc/sys/net/bridge folder is only created in the initial network namespace. 
> This blocks use-cases where users would like to e.g. not do bridge filtering for bridges in a specific 
> network namespace while doing so for bridges located in another network namespace.
> Fix: The patches linked below ensure that the /proc/sys/net/bridge folder is available in each network 
> namespace if the module is loaded and disappears from all network namespaces when the module is unloaded.
> In doing so the patch makes the sysctls:
> bridge-nf-call-arptables
> bridge-nf-call-ip6tables
> bridge-nf-call-iptables
> bridge-nf-filter-pppoe-tagged
> bridge-nf-filter-vlan-tagged
> bridge-nf-pass-vlan-input-dev
> apply per network namespace.
> Regression Potential: None, since this didn't use to work before. Otherwise limited to the br_netfilter module.
> The netfilter rules are afaict already per network namespace so it should be safe for users to specify whether 
> bridge devices inside a network namespace are supposed to go through iptables et al. or not. Also, this can 
> already be done per-bridge by setting an option for each individual bridge via Netlink. It should also be 
> possible to do this for all bridges in a network namespace via sysctls.
> Test Case: Tested with LXD on a kernel with the patches applied and per-network namespace iptables.

I don't agree with the regression potential here, which is particularly
ironic given that the final patch fixes a user-after-free introduced by
the patch before it (I'm sitting next to Christian as I type this and
have already given him a hard time about that one). They are also fairly
new, being upstream as of 5.3-rc1. So I think we need a better statement
of regression potential and what kind of regression testing as been done
before considering them for SRU.

Applied to eoan/master-next.


More information about the kernel-team mailing list