Re: [PATCH v2 bpf-next 02/12] libbpf: implement BPF CO-RE offset relocation algorithm

[Song Liu <songliubraving@fb.com>]

> On Jul 31, 2019, at 10:18 AM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
> 
> On Wed, Jul 31, 2019 at 1:30 AM Song Liu <songliubraving@fb.com> wrote:
>> 
>> 
>> 
>>> On Jul 30, 2019, at 11:52 PM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
>>> 
>>> On Tue, Jul 30, 2019 at 10:19 PM Song Liu <songliubraving@fb.com> wrote:
>>>> 
>>>> 
>>>> 
>>>>> On Jul 30, 2019, at 6:00 PM, Andrii Nakryiko <andrii.nakryiko@gmail.com> wrote:
>>>>> 
>>>>> On Tue, Jul 30, 2019 at 5:39 PM Song Liu <songliubraving@fb.com> wrote:
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>>> On Jul 30, 2019, at 12:53 PM, Andrii Nakryiko <andriin@fb.com> wrote:
>>>>>>> 
>>>>>>> This patch implements the core logic for BPF CO-RE offsets relocations.
>>>>>>> Every instruction that needs to be relocated has corresponding
>>>>>>> bpf_offset_reloc as part of BTF.ext. Relocations are performed by trying
>>>>>>> to match recorded "local" relocation spec against potentially many
>>>>>>> compatible "target" types, creating corresponding spec. Details of the
>>>>>>> algorithm are noted in corresponding comments in the code.
>>>>>>> 
>>>>>>> Signed-off-by: Andrii Nakryiko <andriin@fb.com>
>> 
>> [...]
>> 
>>>>>> 
>>>>> 
>>>>> I just picked the most succinct and non-repetitive form. It's
>>>>> immediately apparent which type it's implicitly converted to, so I
>>>>> felt there is no need to repeat it. Also, just (void *) is much
>>>>> shorter. :)
>>>> 
>>>> _All_ other code in btf.c converts the pointer to the target type.
>>> 
>>> Most in libbpf.c doesn't, though. Also, I try to preserve pointer
>>> constness for uses that don't modify BTF types (pretty much all of
>>> them in libbpf), so it becomes really verbose, despite extremely short
>>> variable names:
>>> 
>>> const struct btf_member *m = (const struct btf_member *)(t + 1);
>> 
>> I don't think being verbose is a big problem here. Overusing
> 
> Problem is too big and strong word to describe this :). It hurts
> readability and will often quite artificially force either wrapping
> the line or unnecessarily splitting declaration and assignment. Void *
> on the other hand is short and usually is in the same line as target
> type declaration, if not, you'll have to find local variable
> declaration to double-check type, if you are unsure.
> 
> Using (void *) + implicit cast to target pointer type is not
> unprecedented in libbpf:
> 
> $ rg ' = \((const )?struct \w+ \*\)' tools/lib/bpf/ | wc -l
> 52
> $ rg ' = \((const )?void \*\)' tools/lib/bpf/  | wc -l
> 35
> 
> 52 vs 35 is majority overall, but not by a landslide.
> 
>> (void *) feels like a bigger problem.
> 
> Why do you feel it's a problem? void * conveys that we have a piece of
> memory that we will need to reinterpret as some concrete pointer type.
> That's what we are doing, skipping btf_type and then interpreting
> memory after common btf_type prefix is some other type, depending on
> actual BTF kind. I don't think void * is misleading in any way.

(void *) hides some problem. For example:

	struct type_a *ptr_a = NULL;
	struct type_b *ptr_b = NULL;

	/* we want this */
	ptr_a = (struct type_a *)data;
	ptr_b = (struct type_b *)(data + offset);

	/* typo, should be ptr_b, compiler will complain */
	ptr_a = (struct type_b *)(data + offset);

	/* typo, should be ptr_b, compiler will ignore */
	ptr_a = (void *)(data + offset);

Such typo is not very rare. And it may be really painful to debug. 

That being said, I think we have spent too much time on this. I will 
let you make the final call. Either way:

Acked-by: Song Liu <songliubraving@fb.com>