Re: [PATCH bpf-next v3 2/2] selftests/bpf: Add ldsx selftests for ldsx and subreg compare

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

On 7/16/24 5:12 PM, Eduard Zingerman wrote:
> On Tue, 2024-07-16 at 15:38 -0700, Yonghong Song wrote:
>
> [...]
>
>> diff --git a/tools/testing/selftests/bpf/prog_tests/reg_bounds.c b/tools/testing/selftests/bpf/prog_tests/reg_bounds.c
>> index eb74363f9f70..c88602908cfe 100644
>> --- a/tools/testing/selftests/bpf/prog_tests/reg_bounds.c
>> +++ b/tools/testing/selftests/bpf/prog_tests/reg_bounds.c
>> @@ -441,6 +441,22 @@ static struct range range_refine(enum num_t x_t, struct range x, enum num_t y_t,
>>           if (t_is_32(y_t) && !t_is_32(x_t)) {
>>                   struct range x_swap;
>>    
>> +               /* If we know that
>> +                *   - *x* is in the range of signed 32bit value
>> +                *   - *y_cast* range is 32-bit sign non-negative, and
>> +                * then *x* range can be narrowed to the interaction of
>> +                * *x* and *y_cast*. Otherwise, if the new range for *x*
>> +                * allows upper 32-bit 0xffffffff then the eventual new
>> +                * range for *x* will be out of signed 32-bit range
>> +                * which violates the origin *x* range.
>> +                */
>> +               if (x_t == S64 && y_t == S32 &&
>> +                   !(y_cast.a & 0xffffffff80000000ULL) && !(y_cast.b & 0xffffffff80000000) &&
>> +                   (long long)x.a >= S32_MIN && (long long)x.b <= S32_MAX) {
>> +                               return range(S64, max_t(S64, y_cast.a, x.a),
>> +                                            min_t(S64, y_cast.b, x.b));
>> +               }
>> +
>>                   /* some combinations of upper 32 bits and sign bit can lead to
>>                    * invalid ranges, in such cases it's easier to detect them
>>                    * after cast/swap than try to enumerate all the conditions
>> @@ -2108,6 +2124,9 @@ static struct subtest_case crafted_cases[] = {
>>           {S32, U32, {(u32)S32_MIN, 0}, {0, 0}},
>>           {S32, U32, {(u32)S32_MIN, 0}, {(u32)S32_MIN, (u32)S32_MIN}},
>>           {S32, U32, {(u32)S32_MIN, S32_MAX}, {S32_MAX, S32_MAX}},
>> +       {S64, U32, {0x0, 0x1f}, {0xffffffff80000000ULL, 0x000000007fffffffULL}},
>> +       {S64, U32, {0x0, 0x1f}, {0xffffffffffff8000ULL, 0x0000000000007fffULL}},
>> +       {S64, U32, {0x0, 0x1f}, {0xffffffffffffff80ULL, 0x000000000000007fULL}},
>>    };
>>    
>>    /* Go over crafted hard-coded cases. This is fast, so we do it as part of
>>
>> The logic is very similar to kernel implementation but has a difference in generating
>> the final range. In reg_bounds implementation, the range is narrowed by intersecting
>> y_cast and x range which are necessary.
>>
>> In kernel implementation, there is no interection since we only have one register
>> and two register has been compared before.
>>
>> Eduard, could you take a look at the above code?
> I think this change is correct.
> The return clause could be simplified a bit:
>
> 	return range_improve(x_t, x, y_cast);

Indeed. This is much simpler. I will use reg_bounds testing instead of verifier_ldsx testing
in next revision.

>
> [...]