socket, bpf: fix sk_filter use after free in sk_clone_lock
authorDaniel Borkmann <daniel@iogearbox.net>
Wed, 22 Mar 2017 12:08:08 +0000 (13:08 +0100)
committerDavid S. Miller <davem@davemloft.net>
Wed, 22 Mar 2017 22:37:04 +0000 (15:37 -0700)
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>
net/core/sock.c

index acb0d41..2c4f574 100644 (file)
@@ -1544,6 +1544,12 @@ struct sock *sk_clone_lock(const struct sock *sk, const gfp_t priority)
                        is_charged = sk_filter_charge(newsk, filter);
 
                if (unlikely(!is_charged || xfrm_sk_clone_policy(newsk, sk))) {
+                       /* We need to make sure that we don't uncharge the new
+                        * socket if we couldn't charge it in the first place
+                        * as otherwise we uncharge the parent's filter.
+                        */
+                       if (!is_charged)
+                               RCU_INIT_POINTER(newsk->sk_filter, NULL);
                        sk_free_unlock_clone(newsk);
                        newsk = NULL;
                        goto out;