l2tp_tunnel_create() inserts the new tunnel into the namespace's tunnel
list and sets the socket's ->sk_user_data field, before returning it to
the caller. Therefore, there are two ways the tunnel can be accessed
and freed, before the caller even had the opportunity to take a
reference. In practice, syzbot could crash the module by closing the
socket right after a new tunnel was returned to pppol2tp_create().

This patch moves tunnel registration out of l2tp_tunnel_create(), so
that the caller can safely hold a reference before publishing the
tunnel. This second step is done with the new l2tp_tunnel_register()
function, which is now responsible for associating the tunnel to its
socket and for inserting it into the namespace's list.

While moving the code to l2tp_tunnel_register(), a few modifications
have been done. First, the socket validation tests are done in a helper
function, for clarity. Also, modifying the socket is now done after
having inserted the tunnel to the namespace's tunnels list. This will
allow insertion to fail, without having to revert theses modifications
in the error path (a followup patch will check for duplicate tunnels
before insertion). Either the socket is a kernel socket which we
control, or it is a user-space socket for which we have a reference on
the file descriptor. In any case, the socket isn't going to be closed
from under us.

Reported-by: syzbot+fbeeb5c3b538e8545644@syzkaller.appspotmail.com
Fixes: fd558d18 ("l2tp: Split pppol2tp patch into separate l2tp and ppp parts")
Signed-off-by: Guillaume Nault <g.nault@alphalink.fr>
Signed-off-by: David S. Miller <davem@davemloft.net>