Re: [PATCH bpf-next] bpf/verifier: enable ctx + const + 0.

[Alexei Starovoitov <alexei.starovoitov@gmail.com>]

On Mon, Apr 30, 2018 at 10:15:05AM -0700, William Tu wrote:
> Existing verifier does not allow 'ctx + const + const'.  However, due to
> compiler optimization, there is a case where BPF compilerit generates
> 'ctx + const + 0', as shown below:
> 
>   599: (1d) if r2 == r4 goto pc+2
>    R0=inv(id=0) R1=ctx(id=0,off=40,imm=0)
>    R2=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff))
>    R3=inv(id=0,umax_value=65535,var_off=(0x0; 0xffff)) R4=inv0
>    R6=ctx(id=0,off=0,imm=0) R7=inv2
>   600: (bf) r1 = r6			// r1 is ctx
>   601: (07) r1 += 36			// r1 has offset 36
>   602: (61) r4 = *(u32 *)(r1 +0)	// r1 + 0
>   dereference of modified ctx ptr R1 off=36+0, ctx+const is allowed,
>   ctx+const+const is not
> 
> The reason for BPF backend generating this code is due optimization
> likes this, explained from Yonghong:
>     if (...)
>         *(ctx + 60)
>     else
>         *(ctx + 56)
> 
> The compiler translates it to
>     if (...)
>        ptr = ctx + 60
>     else
>        ptr = ctx + 56
>     *(ptr + 0)
> 
> So load ptr memory become an example of 'ctx + const + 0'.  This patch
> enables support for this case.
> 
> Fixes: f8ddadc4db6c7 ("Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net")
> Cc: Yonghong Song <yhs@fb.com>
> Signed-off-by: Yifeng Sun <pkusunyifeng@gmail.com>
> Signed-off-by: William Tu <u9012063@gmail.com>
> ---
>  kernel/bpf/verifier.c                       |  2 +-
>  tools/testing/selftests/bpf/test_verifier.c | 13 +++++++++++++
>  2 files changed, 14 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index 712d8655e916..c9a791b9cf2a 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -1638,7 +1638,7 @@ static int check_mem_access(struct bpf_verifier_env *env, int insn_idx, u32 regn
>  		/* ctx accesses must be at a fixed offset, so that we can
>  		 * determine what type of data were returned.
>  		 */
> -		if (reg->off) {
> +		if (reg->off && off != reg->off) {
>  			verbose(env,
>  				"dereference of modified ctx ptr R%d off=%d+%d, ctx+const is allowed, ctx+const+const is not\n",
>  				regno, reg->off, off - reg->off);
> diff --git a/tools/testing/selftests/bpf/test_verifier.c b/tools/testing/selftests/bpf/test_verifier.c
> index 1acafe26498b..95ad5d5723ae 100644
> --- a/tools/testing/selftests/bpf/test_verifier.c
> +++ b/tools/testing/selftests/bpf/test_verifier.c
> @@ -8452,6 +8452,19 @@ static struct bpf_test tests[] = {
>  		.prog_type = BPF_PROG_TYPE_SCHED_CLS,
>  	},
>  	{
> +		"arithmetic ops make PTR_TO_CTX + const + 0 valid",
> +		.insns = {
> +			BPF_ALU64_IMM(BPF_ADD, BPF_REG_1,
> +				      offsetof(struct __sk_buff, data) -
> +				      offsetof(struct __sk_buff, mark)),
> +			BPF_LDX_MEM(BPF_W, BPF_REG_0, BPF_REG_1, 0),

How rewritten code looks here?

The patch is allowing check_ctx_access() to proceed with sort-of
correct 'off' and remember ctx_field_size,
but in convert_ctx_accesses() it's using insn->off to do conversion.
Which is zero in this case, so it will convert
struct __sk_buff {
        __u32 len; // offset 0

into access of 'struct sk_buff'->len
and then will add __sk_buff's &data - &mark delta to in-kernel len field.
Which will point to some random field further down in struct sk_buff.
Doesn't look correct at all.

How did you test this patch?