Re: [PATCH bpf] bpf: Fix a verifier bug due to incorrect branch offset comparison with cpu=v4

[Yonghong Song <yonghong.song@linux.dev>]

On 11/29/23 6:42 PM, Yonghong Song wrote:
>
> On 11/29/23 5:51 PM, Daniel Borkmann wrote:
>> On 11/29/23 8:54 AM, Yonghong Song wrote:
>>> Bpf cpu=v4 support is introduced in [1] and Commit 4cd58e9af8b9
>>> ("bpf: Support new 32bit offset jmp instruction") added support for new
>>> 32bit offset jmp instruction. Unfortunately, in function
>>> bpf_adj_delta_to_off(), for new branch insn with 32bit offset, the 
>>> offset
>>> (plus/minor a small delta) compares to 16-bit offset bound
>>> [S16_MIN, S16_MAX], which caused the following verification failure:
>>>    $ ./test_progs-cpuv4 -t verif_scale_pyperf180
>>>    ...
>>>    insn 10 cannot be patched due to 16-bit range
>>>    ...
>>>    libbpf: failed to load object 'pyperf180.bpf.o'
>>>    scale_test:FAIL:expect_success unexpected error: -12 (errno 12)
>>>    #405     verif_scale_pyperf180:FAIL
>>>
>>> Note that due to recent llvm18 development, the patch [2] (already 
>>> applied
>>> in bpf-next) needs to be applied to bpf tree for testing purpose.
>>>
>>> The fix is rather simple. For 32bit offset branch insn, the adjusted
>>> offset compares to [S32_MIN, S32_MAX] and then verification succeeded.
>>>
>>>    [1] 
>>> https://lore.kernel.org/all/20230728011143.3710005-1-yonghong.song@linux.dev
>>>    [2] 
>>> https://lore.kernel.org/bpf/20231110193644.3130906-1-yonghong.song@linux.dev
>>>
>>> Fixes: 4cd58e9af8b9 ("bpf: Support new 32bit offset jmp instruction")
>>> Signed-off-by: Yonghong Song <yonghong.song@linux.dev>
>>> ---
>>>   kernel/bpf/core.c | 9 ++++++---
>>>   1 file changed, 6 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
>>> index cd3afe57ece3..74f2fd48148c 100644
>>> --- a/kernel/bpf/core.c
>>> +++ b/kernel/bpf/core.c
>>> @@ -371,14 +371,17 @@ static int bpf_adj_delta_to_imm(struct 
>>> bpf_insn *insn, u32 pos, s32 end_old,
>>>   static int bpf_adj_delta_to_off(struct bpf_insn *insn, u32 pos, 
>>> s32 end_old,
>>>                   s32 end_new, s32 curr, const bool probe_pass)
>>>   {
>>> -    const s32 off_min = S16_MIN, off_max = S16_MAX;
>>> +    s32 off_min = S16_MIN, off_max = S16_MAX;
>>>       s32 delta = end_new - end_old;
>>>       s32 off;
>>
>> These should all be converted to s64, no? E.g. further below
>> the test will never trigger then for jmp32:
>>
>>        if (off < off_min || off > off_max)
>>                 return -ERANGE;
>
>
> good point! Let us use s64 for potential overflows.
> Will send v2 soon.

I didn't change 's32 delta' type to be consistent with
bpf_adj_delta_to_imm() such that the delta should be
within s32 range. Technically off_min/off_max can
remain as 's32' but I changed them to 's64' to be consistent
with bpf_adj_delta_to_imm().

>
>>
>>> -    if (insn->code == (BPF_JMP32 | BPF_JA))
>>> +    if (insn->code == (BPF_JMP32 | BPF_JA)) {
>>>           off = insn->imm;
>>> -    else
>>> +        off_min = S32_MIN;
>>> +        off_max = S32_MAX;
>>> +    } else {
>>>           off = insn->off;
>>> +    }
>>>         if (curr < pos && curr + off + 1 >= end_old)
>>>           off += delta;
>>>
>>
>