Re: [PATCH RFC bpf-next 1/4] bpf: add struct largest member size in func model

["Alexis Lothoré" <alexis.lothore@bootlin.com>]

Hello Alexei,

On Fri Apr 25, 2025 at 1:14 AM CEST, Alexei Starovoitov wrote:
> On Thu, Apr 24, 2025 at 6:38 AM Alexis Lothoré
> <alexis.lothore@bootlin.com> wrote:

[...]

>> > With DWARF info, we might not need to detect the structure alignment anymore,
>> > since the DW_AT_location attribute tells us where the structure parameter is
>> > located on the stack, and DW_AT_byte_size gives us the size of the structure.
>>
>> I am not sure to follow you here, because DWARF info is not accessible
>> from kernel at runtime, right ? Or are you meaning that we could, at build
>> time, enrich the BTF info embedded in the kernel thanks to DWARF info ?
>
> Sounds like arm64 has complicated rules for stack alignment and
> stack offset computation for passing 9th+ argument.

AFAICT, arm64 has some specificities for large types, but not that much
compared to x86 for example. If I take a look at System V ABI ([1]), I see
pretty much the same constraints:
- p.18: "Arguments of type __int128 offer the same operations as INTEGERs,
  [...] with the exception that arguments of type __int128 that are stored
  in memory must be aligned on a 16-byte boundary"
- p.13: "Structures and unions assume the alignment of their most strictly
  aligned component"
- the custom packing and alignments attributes will end up having the same
  consequence on both architectures

As I mentioned in my cover letter, the new tests covering those same
alignment constraints for ARM64 break on x86, which makes me think other
archs are also silently ignoring those cases.

> Since your analysis shows:
> "there are about 200 functions accept 9 to 12 arguments, so adding support
> for up to 12 function arguments."
> I say, let's keep the existing limitation:
>         if (nregs > 8)
>                 return -ENOTSUPP;
>
> If there is a simple and dumb way to detect that arg9+ are scalars
> with simple stack passing rules, then, sure, let's support those too,
> but fancy packed/align(x)/etc let's ignore.


[1] https://refspecs.linuxbase.org/elf/x86_64-abi-0.99.pdf


-- 
Alexis Lothoré, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com