bpf: Fix clobbering of r2 in bpf_gen_ld_abs

Bryce reported that he saw the following with:

  0:  r6 = r1
  1:  r1 = 12
  2:  r0 = *(u16 *)skb[r1]

The xlated sequence was incorrectly clobbering r2 with pointer
value of r6 ...

  0: (bf) r6 = r1
  1: (b7) r1 = 12
  2: (bf) r1 = r6
  3: (bf) r2 = r1
  4: (85) call bpf_skb_load_helper_16_no_cache#7692160

... and hence call to the load helper never succeeded given the
offset was too high. Fix it by reordering the load of r6 to r1.

Other than that the insn has similar calling convention than BPF
helpers, that is, r0 - r5 are scratch regs, so nothing else
affected after the insn.

Fixes: e0cea7c ("bpf: implement ld_abs/ld_ind in native bpf")
Reported-by: Bryce Kahle <bryce.kahle@datadoghq.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Link: https://lore.kernel.org/bpf/cace836e4d07bb63b1a53e49c5dfb238a040c298.1599512096.git.daniel@iogearbox.net

net/core/filter.c

@@ -7065,15 +7065,15 @@ static int bpf_gen_ld_abs(const struct bpf_insn *orig,
 	bool indirect = BPF_MODE(orig->code) == BPF_IND;
 	struct bpf_insn *insn = insn_buf;
 
-	/* We're guaranteed here that CTX is in R6. */
-	*insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_CTX);
 	if (!indirect) {
 		*insn++ = BPF_MOV64_IMM(BPF_REG_2, orig->imm);
 	} else {
 		*insn++ = BPF_MOV64_REG(BPF_REG_2, orig->src_reg);
 		if (orig->imm)
 			*insn++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_2, orig->imm);
 	}
+	/* We're guaranteed here that CTX is in R6. */
+	*insn++ = BPF_MOV64_REG(BPF_REG_1, BPF_REG_CTX);
 
 	switch (BPF_SIZE(orig->code)) {
 	case BPF_B: