r333175 converted the global multicast lock to a sleepable sx lock, so
the lock order wrt the (non-sleepable) inp lock changed. To handle this,
r333175 and r333505 introduced a pattern of
However, this means that inp_join_group() and inp_leave_group() aren't
properly serialized against each other. Concurrent attempts to add a
souce address to a group and leave a group can leave things in an
inconsistent state and trigger panics. Also, the handling for the case
where we release the last inp ref was broken.
Fix the problem by simply acquiring the global multicast lock earlier.
I also fixed some lingering LORs that were not covered by r333505. I am
not convinced that the global lock is actually required anymore
(especially since it is used differently between IPv4 and IPv6), but I
need to study the code further. In the meantime, the change fixes
several races, present in 12.0, found by syzkaller.