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

[Dan Li <ashimida.1990@gmail.com>]

Hi Mark,

Sorry for the late reply.

On 01/03, Mark Rutland wrote:
> On Sun, Dec 18, 2022 at 10:17:58PM -0800, Dan Li wrote:
> >    If there are user mode programs or other systems that want to use
> >    this feature, it may be more convenient to use a callback (so this
> >    compilation option is set to -fsanitize=cfi instead of kcfi).
> 
> I appreciate that may be nicer for userspace, but it would be far nicer for the
> kernel if we could have a kcfi mode that behaves the same as LLVM, using a BRK.
> That's going to be simpler for the kernel to deal with, and should result in
> nicer code / smaller binary size (for the reasons given above).
> 
> Can we please have an LLVM-compatible KCFI mode, and have the -fsanitize=cfi be
> a separate option from -fsanitize=kcfi?

Ok, in the next version I will change to the same option as clang :)

> 
> > 2. A reserved typeid (such as 0x0U on the aarch64 platform) is always
> >    inserted in front of functions that should not be called indirectly.
> >    Functions that can be called indirectly will not use this hash value,
> >    which prevents instructions/data before the function from being used
> >    as a typeid by an attacker.
> 
> That sounds sensible, though it meanse we'll need to go audit all the assembly
> without type annotations.
> 
> I presume that "functions that should not be called indirectly" only includes
> those which are not directly visible outside the compilation unit AND whose
> address is never taken / escaped from the compilation unit. Is that the case?

Yes.

> 
> > 3. Some bits are ignored in the typeid to avoid conflicts between the
> >    typeid and the instruction set of a specific platform, thereby
> >    preventing an attacker from bypassing the CFI check by using the
> >    instruction as a typeid, such as on the aarch64 platform:
> >    * If the following instruction sequence exists:
> > 	  400620:       a9be7bfd        stp     x29, x30, [sp, #-32]!
> > 	  400624:       910003fd        mov     x29, sp
> > 	  400628:       f9000bf3        str     x19, [sp, #16]
> >    * If the expected typeid of the indirect call is exactly 0x910003fd,
> >      the attacker can jump to the next instruction position of any
> >      "mov x29,sp" instruction (such as 0x400628 here).
> 
> Which bits exactly are ignored on arm64?
> 
> e.g. are these encoded into UDF immediates?

In aarch64, I currently ignore bit [28:27]. IUCC, according to the manual[1],
it is a UDF instruction only when the upper 16 bits are all 0.
But due to this has too much impact on the entropy of typeid, so I (not
rigorously) only ignore 2 bits here, and most of the instruction codes covered
by it belong to 'Reserved' or 'UNALLOCATED' (probably not a good idea).

But as Kees said, if clang doesn't handle it here, in order to be consistent,
I think it's better for gcc to not handle it when implementing kernel cfi.

[1] https://developer.arm.com/documentation/ddi0602/2022-06/Index-by-Encoding?lang=en

> 
> As a general thing, how does this work with -fpatchable-function-entry=M,N,
> where N is non-zero?
> 
> We still need to fix that for LLVM, and it would be good to align on the same behaviour.
>

Yeah, it makes sense.

Currently, it is consistent with llvm. Taking -fpatchable-function-entry=2,1
as an example, the currently generated code is as follows:

__cfi_main:
        .4byte 0x439d3502
        .global main
        .section        __patchable_function_entries
        .align  3
        .8byte  .LPFE3
        .text
.LPFE3:
        nop
        .type   main, %function
main:
        nop
.LFB2:
        .cfi_startproc
        stp     x29, x30, [sp, -32]!

Finally, do we want to generate code like this?
        nop
        .4byte 0x439d3502
main:
        nop
        ...

Thanks,
Dan.

> >  
> > -- 
> > 2.17.1
> >