Re: [PATCH bpf-next v5 7/9] selftests/bpf: Test kptr arrays and kptrs in nested struct fields.

[Kui-Feng Lee <sinquersw@gmail.com>]

On 5/10/24 15:08, Eduard Zingerman wrote:
> On Fri, 2024-05-10 at 14:59 -0700, Kui-Feng Lee wrote:
>>
>> For the sake of completeness, would it be possible to create a test
>>> case where there are several struct arrays following each other?
>>> E.g. as below:
>>>
>>> struct foo {
>>>     ... __kptr *a;
>>>     ... __kptr *b;
>>> }
>>>
>>> struct bar {
>>>     ... __kptr *c;
>>> }
>>>
>>> struct {
>>>     struct foo foos[3];
>>>     struct bar bars[2];
>>> }
>>>
>>> Just to check that offset is propagated correctly.
>>
>> Sure!
> 
> Great, thank you
> 
>>> Also, in the tests below you check that a pointer to some object could
>>> be put into an array at different indexes. Tbh, I find it not very
>>> interesting if we want to check that offsets are correct.
>>> Would it be possible to create an array of object kptrs,
>>> put specific references at specific indexes and somehow check which
>>> object ended up where? (not necessarily 'bpf_cpumask').
>>
>> Do you mean checking index in the way like the following code?
>>
>>    if (array[0] != ref0 || array[1] != ref1 || array[2] != ref2 ....)
>>      return err;
> 
> Probably, but I'd need your help here.
> There goal is to verify that offsets of __kptr's in the 'info' array
> had been set correctly. Where is this information is used later on?
> E.g. I'd like to trigger some action that "touches" __kptr at index N
> and verify that all others had not been "touched".
> But this "touch" action has to use offset stored in the 'info'.


They are used for verifying the offset of instructions.
Let's assume we have an array of size 10.
Then, we have 10 infos with 10 different offsets.
And, we have a program includes one instruction for each element, 10 in
total, to access the corresponding element.
Each instruction has an offset different from others, generated by the 
compiler. That means the verifier will fail to find an info for some of
instructions if there is one or more info having wrong offset.



> 
> [...]