The old approach where we go through the list of interfaces and count them
has bugs. One obvious bug with this dynamic translation is that once an
Ethernet interface in the middle of the list goes away, all interfaces
following it would change their Linux names.
A bigger problem is the ifnet arrival and departure times. For example
linsysfs has event handler for ifnet_arrival_event, and of course it wants
to resolve the name. This accidentially works, due to a bug in
if_attach() where we call if_link_ifnet() before invoking all the event
handlers. Once the bug is fixed linsysfs won't be able to resolve the old
way. The other side is ifnet_departure_event, where there is no bug, the
eventhandlers are called after the if_unlink_ifnet(). This means old
translation won't work for departure event handlers. One example is
netlink. This change gives the Netlink a chance to emit a proper Linux
interface departure message.
However, there is another problem in Netlink, that the ifnet pointer is
lost in the Netlink translation layer. Plug this with a cookie in netlink
writer structure that can be set by the route layer and used by the Netlink
Linux translation layer. This part of the diff seems unrelated, but it is
hard to make it a separate change, as the old KPI goes away and to use the
new one we need the pointer.
Differential Revision: https://reviews.freebsd.org/D54077