bpf: Avoid verifier failure for 32bit pointer arithmetic

When do experiments with llvm (disabling instcombine and
simplifyCFG), I hit the following error with test_seg6_loop.o.

  ; R1=pkt(id=0,off=0,r=48,imm=0), R7=pkt(id=0,off=40,r=48,imm=0)
  w2 = w7
  ; R2_w=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff))
  w2 -= w1
  R2 32-bit pointer arithmetic prohibited

The corresponding source code is:
  uint32_t srh_off
  // srh and skb->data are all packet pointers
  srh_off = (char *)srh - (char *)(long)skb->data;

The verifier does not support 32-bit pointer/scalar arithmetic.

Without my llvm change, the code looks like

  ; R3=pkt(id=0,off=40,r=48,imm=0), R8=pkt(id=0,off=0,r=48,imm=0)
  w3 -= w8
  ; R3_w=inv(id=0)

This is explicitly allowed in verifier if both registers are
pointers and the opcode is BPF_SUB.

To fix this problem, I changed the verifier to allow
32-bit pointer/scaler BPF_SUB operations.

At the source level, the issue could be workarounded with
inline asm or changing "uint32_t srh_off" to "uint64_t srh_off".
But I feel that verifier change might be the right thing to do.
Signed-off-by: NYonghong Song <yhs@fb.com>
Signed-off-by: NDaniel Borkmann <daniel@iogearbox.net>
Acked-by: NJohn Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20200618234631.3321118-1-yhs@fb.com

kernel/bpf/verifier.c

@@ -5031,6 +5031,11 @@ static int adjust_ptr_min_max_vals(struct bpf_verifier_env *env,
 
 	if (BPF_CLASS(insn->code) != BPF_ALU64) {
 		/* 32-bit ALU ops on pointers produce (meaningless) scalars */
+		if (opcode == BPF_SUB && env->allow_ptr_leaks) {
+			__mark_reg_unknown(env, dst_reg);
+			return 0;
+		}
+
 		verbose(env,
 			"R%d 32-bit pointer arithmetic prohibited\n",
 			dst);