In sk_clone_lock(), we create a new socket and inherit most of the
parent's members via sock_copy() which memcpy()'s various sections.
Now, in case the parent socket had a BPF socket filter attached,
then newsk->sk_filter points to the same instance as the original
sk->sk_filter.

sk_filter_charge() is then called on the newsk->sk_filter to take a
reference and should that fail due to hitting max optmem, we bail
out and release the newsk instance.

The issue is that commit 278571baca2a ("net: filter: simplify socket
charging") wrongly combined the dismantle path with the failure path
of xfrm_sk_clone_policy(). This means, even when charging failed, we
call sk_free_unlock_clone() on the newsk, which then still points to
the same sk_filter as the original sk.

Thus, sk_free_unlock_clone() calls into __sk_destruct() eventually
where it tests for present sk_filter and calls sk_filter_uncharge()
on it, which potentially lets sk_omem_alloc wrap around and releases
the eBPF prog and sk_filter structure from the (still intact) parent.

Fix it by making sure that when sk_filter_charge() failed, we reset
newsk->sk_filter back to NULL before passing to sk_free_unlock_clone(),
so that we don't mess with the parents sk_filter.

Only if xfrm_sk_clone_policy() fails, we did reach the point where
either the parent's filter was NULL and as a result newsk's as well
or where we previously had a successful sk_filter_charge(), thus for
that case, we do need sk_filter_uncharge() to release the prior taken
reference on sk_filter.

Fixes: 278571baca2a ("net: filter: simplify socket charging")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>