Linux neighbour tables become per-network-namespace

Kuniyuki Iwashima has submitted work for net-next that turns the kernel's global ARP and IPv6 neighbour discovery tables into per-network-namespace tables. The neighbour subsystem was already close to dropping the RTNL for control paths, but those paths still contended on a single per-table lock. Namespacing the tables is the step that breaks that global choke point.

The practical user-visible win is sysctl. Paths under /proc/sys/net/ipv4/neigh/default/ and the IPv6 counterpart were effectively global: only init_net could set them. After this change, each netns gets its own defaults and garbage-collection parameters, and RTM_SETNEIGHTBL can adjust the default table and GC knobs per namespace.

To limit surprise for existing setups, new namespaces inherit those defaults from init_net. A new core sysctl, net.core.neigh_inherit_init_net, controls that. It is on by default; turning it off gives a fresh netns the stock neighbour parameters instead of whatever init_net was using.

The conversion also cleans up leftover global assumptions: proxy neighbour entries no longer carry a net pointer, several obsolete exports and RCU protections around the old global table list go away, and table lifetime is tied to a refcount so a namespace can tear down without racing neighbour timers. A few drivers that still touch the IPv6 table without checking whether IPv6 is disabled keep a compatibility path for now.

For operators running many netns or containers, neighbour GC thresholds and related tunables stop being a machine-wide compromise. For the networking stack, the series is another concrete move toward lighter locking on neighbour add and delete.