bpf: cgroup skb progs cannot access ld_abs/ind

Commit fb9a307 ("bpf: Allow CGROUP_SKB eBPF program to
access sk_buff") enabled programs of BPF_PROG_TYPE_CGROUP_SKB
type to use ld_abs/ind instructions. However, at this point,
we cannot use them, since offsets relative to SKF_LL_OFF will
end up pointing skb_mac_header(skb) out of bounds since in the
egress path it is not yet set at that point in time, but only
after __dev_queue_xmit() did a general reset on the mac header.
bpf_internal_load_pointer_neg_helper() will then end up reading
data from a wrong offset.

BPF_PROG_TYPE_CGROUP_SKB programs can use bpf_skb_load_bytes()
already to access packet data, which is also more flexible than
the insns carried over from cBPF.

Fixes: fb9a307 ("bpf: Allow CGROUP_SKB eBPF program to access sk_buff")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Cc: Chenbo Feng <fengc@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

kernel/bpf/verifier.c

@@ -2426,7 +2426,6 @@ static bool may_access_skb(enum bpf_prog_type type)
 	case BPF_PROG_TYPE_SOCKET_FILTER:
 	case BPF_PROG_TYPE_SCHED_CLS:
 	case BPF_PROG_TYPE_SCHED_ACT:
-	case BPF_PROG_TYPE_CGROUP_SKB:
 		return true;
 	default:
 		return false;