Re: [RFC/RFT,V2] CFI: Add support for gcc CFI in aarch64

[Mark Rutland]

On Wed, Dec 13, 2023 at 05:01:07PM +0800, Wang wrote:
> On 2023/12/13 16:48, Dan Li wrote:
> > + Likun
> >
> > On Tue, 28 Mar 2023 at 06:18, Sami Tolvanen wrote:
> >> On Mon, Mar 27, 2023 at 2:30 AM Peter Zijlstra wrote:
> >>> On Sat, Mar 25, 2023 at 01:54:16AM -0700, Dan Li wrote:
> >>>
> >>>> In the compiler part[4], most of the content is the same as Sami's
> >>>> implementation[3], except for some minor differences, mainly including:
> >>>>
> >>>> 1. The function typeid is calculated differently and it is difficult
> >>>> to be consistent.
> >>> This means there is an effective ABI break between the compilers, which
> >>> is sad :-( Is there really nothing to be done about this?
> >> I agree, this would be unfortunate, and would also be a compatibility
> >> issue with rustc where there's ongoing work to support
> >> clang-compatible CFI type hashes:
> >>
> >> https://github.com/rust-lang/rust/pull/105452
> >>
> >> Sami
> 
> Hi Peter and Sami
> 
> I am Dan Li's colleague, and I will take over and continue the work of CFI.
> 
> Regarding the issue of gcc cfi type id being compatible with clang, we
> have analyzed and verified:
> 
> 1. clang uses Mangling defined in Itanium C++ ABI to encode the function
> prototype, and uses the encoding result as input to generate cfi type id;
> 2. Currently, gcc only implements mangling for the C++ compiler, and the
> function prototype coding generated by these interfaces is compatible
> with clang, but gcc's c compiler does not support mangling.;
> 
> Adding mangling to gcc's c compiler is a huge and difficult task，because
> we have to refactor the mangling of C++, splitting it into basic
> mangling and language specific mangling, and adding support for the c
> language which requires a deep understanding of the compiler and
> language processing parts.
> 
> And for the kernel cfi, I suggest separating type compatibility from CFI
> basic functions. Type compatibility is independent from CFI basic
> funcitons and should be dealt with under another topic. Should we focus
> on the main issus of cfi, and  let it work first on linux kernel, and
> left the compatible issue to be solved later?

I'm not sure what you're suggesting here exactly, do you mean to add a type ID
scheme that's incompatible with clang, leaving everything else the same? If so,
what sort of scheme are you proposing?

It seems unfortunate to have a different scheme, but IIUC we expect all kernel
objects to be built with the same compiler.

Mark.