Re: [PATCH bpf-next 1/2] bpf: update 32-bit bounds when the lower 32-bit value is not wrapping

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, Mar 7, 2023 at 1:05 AM Xu Kuohai <xukuohai@xxxxxxxxxxxxxxx> wrote:
>
> The following XDP prog is accepted by verifier.
>
> 0: (61) r2 = *(u32 *)(r1 +0)          ; R2_w=pkt(off=0,r=0,imm=0)
> 1: (61) r3 = *(u32 *)(r1 +4)          ; R3_w=pkt_end(off=0,imm=0)
> 2: (bf) r1 = r2
> 3: (07) r1 += 1
> 4: (2d) if r1 > r3 goto pc+6
> 5: (71) r1 = *(u8 *)(r2 +0)           ; R1_w=scalar(umax=255,var_off=(0x0; 0xff))
> 6: (b4) w0 = 0x7fffff10
> 7: (0c) w1 += w0                      ; R1_w=scalar(umin=0x7fffff10,umax=0x8000000f,var_off=(0x0; 0xffffffff))
> 8: (b4) w0 = 0x80000000
> 9: (04) w0 += 1
> 10: (ae) if w0 < w1 goto pc-2
> 11: (b7) r0 = 0
> 12: (95) exit
>
> while the following 64-bit version is rejected.
>
> 0: (61) r2 = *(u32 *)(r1 +0)          ; R2_w=pkt(off=0,r=0,imm=0)
> 1: (61) r3 = *(u32 *)(r1 +4)          ; R3_w=pkt_end(off=0,imm=0)
> 2: (bf) r1 = r2
> 3: (07) r1 += 1
> 4: (2d) if r1 > r3 goto pc+8
> 5: (71) r1 = *(u8 *)(r2 +0)           ; R1_w=scalar(umax=255,var_off=(0x0; 0xff))
> 6: (18) r0 = 0x7fffffffffffff10
> 8: (0f) r1 += r0                      ; R1_w=scalar(umin=0x7fffffffffffff10,umax=0x800000000000000f)
> 9: (18) r0 = 0x8000000000000000
> 11: (07) r0 += 1
> 12: (ad) if r0 < r1 goto pc-2
> 13: (b7) r0 = 0
> 14: (95) exit

These two programs are not equivalent.
Not clear how apples to oranges comparison explains anything.

> The verifier log says:
>
> [...]
>
> from 12 to 11: R0_w=-9223372036854775794 R1=scalar(umin=9223372036854775823,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))
> 11: (07) r0 += 1                      ; R0_w=-9223372036854775793
> 12: (ad) if r0 < r1 goto pc-2         ; R0_w=-9223372036854775793 R1=scalar(umin=9223372036854775823,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))
> 13: safe
>
> from 12 to 11: R0_w=-9223372036854775793 R1=scalar(umin=9223372036854775824,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))

First thing to debug is why umin is higher than umax.

> 11: (07) r0 += 1                      ; R0_w=-9223372036854775792
> 12: (ad) if r0 < r1 goto pc-2         ; R0_w=-9223372036854775792 R1=scalar(umin=9223372036854775824,umax=9223372036854775823,var_off=(0x8000000000000000; 0xffffffff))
> 13: safe
>
> [...]
>
> The loop crosses termination condition r0 == r1.umax, and does not stop.
>
> The reason is that when the verifier enumerates to r1.umin == r1.umax, the value
> 0x800000000000000f of r1.umin is greater than U32_MAX, so __reg_combine_64_into_32
> sets the u32 range of r1 to [0, U32_MAX] instead of marking r1 as a constant,
> making is_branch_taken() in check_cond_jmp_op() be skipped.

And it's fine. The verifier is conservative.

>
> To fix it, update 32-bit bounds when the lower 32-bit value is not wrapping,
> even if the 64-bit value is beyond the range of [0, U32_MAX] or [S32_MIN, S32_MAX].

That's not safe in general.

>
> Signed-off-by: Xu Kuohai <xukuohai@xxxxxxxxxxxxxxx>
> ---
>  kernel/bpf/verifier.c | 27 +++++++++++----------------
>  1 file changed, 11 insertions(+), 16 deletions(-)
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index b2116ca78d9a..64c9ee3857ec 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -2013,26 +2013,21 @@ static void __reg_combine_32_into_64(struct bpf_reg_state *reg)
>         reg_bounds_sync(reg);
>  }
>
> -static bool __reg64_bound_s32(s64 a)
> -{
> -       return a >= S32_MIN && a <= S32_MAX;
> -}
> -
> -static bool __reg64_bound_u32(u64 a)
> -{
> -       return a >= U32_MIN && a <= U32_MAX;
> -}
> -
>  static void __reg_combine_64_into_32(struct bpf_reg_state *reg)
>  {
> +       s64 smin = reg->smin_value;
> +       s64 smax = reg->smax_value;
> +       u64 umin = reg->umin_value;
> +       u64 umax = reg->umax_value;
> +
>         __mark_reg32_unbounded(reg);
> -       if (__reg64_bound_s32(reg->smin_value) && __reg64_bound_s32(reg->smax_value)) {
> -               reg->s32_min_value = (s32)reg->smin_value;
> -               reg->s32_max_value = (s32)reg->smax_value;
> +       if ((u64)(smax - smin) <= (u64)U32_MAX && (s32)smin <= (s32)smax) {
> +               reg->s32_min_value = (s32)smin;
> +               reg->s32_max_value = (s32)smax;
>         }
> -       if (__reg64_bound_u32(reg->umin_value) && __reg64_bound_u32(reg->umax_value)) {
> -               reg->u32_min_value = (u32)reg->umin_value;
> -               reg->u32_max_value = (u32)reg->umax_value;
> +       if (umax - umin <= U32_MAX && (u32)umin <= (u32)umax) {
> +               reg->u32_min_value = (u32)umin;
> +               reg->u32_max_value = (u32)umax;

This looks like a workaround for umin > umax issue.
Please debug that instead.

>         }
>         reg_bounds_sync(reg);
>  }
> --
> 2.30.2
>




[Index of Archives]     [Linux Samsung SoC]     [Linux Rockchip SoC]     [Linux Actions SoC]     [Linux for Synopsys ARC Processors]     [Linux NFS]     [Linux NILFS]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]


  Powered by Linux