Re: [PATCH bpf-next 1/2] bpf: Get better reg range with ldsx and 32bit compare

[Eduard Zingerman <eddyz87@gmail.com>]

On Tue, 2024-07-09 at 21:29 -0700, Yonghong Song wrote:

[...]

>   14: (81) r1 = *(s32 *)(r0 +0)         ; R0=rdonly_mem(id=3,ref_obj_id=2,sz=4) R1_w=scalar(smin=0xffffffff80000000,smax=0x7fffffff) refs=2
>   15: (ae) if w1 < w6 goto pc+4 20: R0=rdonly_mem(id=3,ref_obj_id=2,sz=4) R1=scalar(smin=0xffffffff80000000,smax=smax32=umax32=31,umax=0xffffffff0000001f,smin32=0,var_off=(0x0; 0xffffffff0000001f)) R6=scalar(id=1,smin=umin=smin32=umin32=1,smax=umax=smax32=umax32=32,var_off=(0x0; 0x3f)) R7=0 R8=fp-8 R10=fp0 fp-8=iter_num(ref_id=2,state=active,depth=1) refs=2

[...]

> The insn #14 is a sign-extenstion load which is related to 'int i'.
> The insn #15 did a subreg comparision. Note that smin=0xffffffff80000000 and this caused later
> insn #23 failed verification due to unbounded min value.
> 
> Actually insn #15 R1 smin range can be better. Before insn #15, we have
>   R1_w=scalar(smin=0xffffffff80000000,smax=0x7fffffff)
> With the above range, we know for R1, upper 32bit can only be 0xffffffff or 0.
> Otherwise, the value range for R1 could be beyond [smin=0xffffffff80000000,smax=0x7fffffff].
> 
> After insn #15, for the true patch, we know smin32=0 and smax32=32. With the upper 32bit 0xffffffff,
> then the corresponding value is [0xffffffff00000000, 0xffffffff00000020]. The range is
> obviously beyond the original range [smin=0xffffffff80000000,smax=0x7fffffff] and the
> range is not possible. So the upper 32bit must be 0, which implies smin = smin32 and
> smax = smax32.
> 
> This patch fixed the issue by adding additional register deduction after 32-bit compare
> insn such that if the signed 32-bit register range is non-negative and 64-bit smin is
> {S32/S16/S8}_MIN and 64-bit max is no greater than {U32/U16/U8}_MAX.
> Here, we check smin with {S32/S16/S8}_MIN since this is the most common result related to
> signed extension load.

[...]

> Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
> ---
>  kernel/bpf/verifier.c | 15 +++++++++++++++
>  1 file changed, 15 insertions(+)
> 
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index c0263fb5ca4b..3fc557f99b24 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -2182,6 +2182,21 @@ static void __reg_deduce_mixed_bounds(struct bpf_reg_state *reg)
>  		reg->smin_value = max_t(s64, reg->smin_value, new_smin);
>  		reg->smax_value = min_t(s64, reg->smax_value, new_smax);
>  	}
> +
> +	/* if s32 range is non-negative and s64 range is in [S32/S16/S8_MIN, <= S32/S16/S8_MAX],
> +	 * the s64/u64 range can be refined.
> +	 */

Hi Yonghong,

Sorry for delayed response, nice patch, it finally clicked for me.
I'd suggest a slightly different comment, maybe it's just me being
slow, but it took a while to understand why is this correct.
How about a text like below:

  Here we would like to handle a special case after sign extending load,
  when upper bits for a 64-bit range are all 1s or all 0s.

  Upper bits are all 1s when register is in a rage:
    [0xffff_ffff_0000_0000, 0xffff_ffff_ffff_ffff]
  Upper bits are all 0s when register is in a range:
    [0x0000_0000_0000_0000, 0x0000_0000_ffff_ffff]
  Together this forms are continuous range:
    [0xffff_ffff_0000_0000, 0x0000_0000_ffff_ffff]

  Now, suppose that register range is in fact tighter:
    [0xffff_ffff_8000_0000, 0x0000_0000_ffff_ffff] (R)
  Also suppose that it's 32-bit range is positive,
  meaning that lower 32-bits of the full 64-bit register
  are in the range:
    [0x0000_0000, 0x7fff_ffff] (W)

  It so happens, that any value in a range:
    [0xffff_ffff_0000_0000, 0xffff_ffff_7fff_ffff]
  is smaller than a lowest bound of the range (R):
     0xffff_ffff_8000_0000
  which means that upper bits of the full 64-bit register
  can't be all 1s, when lower bits are in range (W).

  Note that:
  - 0xffff_ffff_8000_0000 == (s64)S32_MIN
  - 0x0000_0000_ffff_ffff == (s64)S32_MAX
  These relations are used in the conditions below.

> +	if (reg->s32_min_value >= 0) {
> +		if ((reg->smin_value == S32_MIN && reg->smax_value <= S32_MAX) ||
> +		    (reg->smin_value == S16_MIN && reg->smax_value <= S16_MAX) ||
> +		    (reg->smin_value == S8_MIN && reg->smax_value <= S8_MAX)) {

The explanation above also lands a question, would it be correct to
replace the checks above by a single one?

  reg->smin_value >= S32_MIN && reg->smax_value <= S32_MAX

> +			reg->smin_value = reg->umin_value = reg->s32_min_value;
> +			reg->smax_value = reg->umax_value = reg->s32_max_value;
> +			reg->var_off = tnum_intersect(reg->var_off,
> +						      tnum_range(reg->smin_value,
> +								 reg->smax_value));
> +		}
> +	}
>  }
>  
>  static void __reg_deduce_bounds(struct bpf_reg_state *reg)