Re: [PATCH 05/15] x86: Implement function_nocfi

[Andy Lutomirski]

> On Apr 17, 2021, at 7:20 AM, David Laight <David.Laight@xxxxxxxxxx> wrote:
> 
> From: Kees Cook
>> Sent: 16 April 2021 23:28
>> 
>>> On Fri, Apr 16, 2021 at 03:06:17PM -0700, Andy Lutomirski wrote:
>>> On Fri, Apr 16, 2021 at 3:03 PM Borislav Petkov <bp@xxxxxxxxx> wrote:
>>>> 
>>>> On Fri, Apr 16, 2021 at 02:49:23PM -0700, Sami Tolvanen wrote:
>>>>> __nocfi only disables CFI checking in a function, the compiler still
>>>>> changes function addresses to point to the CFI jump table, which is
>>>>> why we need function_nocfi().
>>>> 
>>>> So call it __func_addr() or get_function_addr() or so, so that at least
>>>> it is clear what this does.
>>>> 
>>> 
>>> This seems backwards to me.  If I do:
>>> 
>>> extern void foo(some signature);
>>> 
>>> then I would, perhaps naively, expect foo to be the actual symbol that
>>> gets called
>> 
>> Yes.
>> 
>>> and for the ABI to be changed to do the CFI checks.
>> 
>> Uh, no? There's no ABI change -- indirect calls are changed to do the
>> checking.
>> 
>>> The
>>> foo symbol would point to whatever magic is needed.
>> 
>> No, the symbol points to the jump table entry. Direct calls get minimal
>> overhead and indirect calls can add the "is this function in the right
>> table" checking.
> 
> 
> Isn't this a bit like one of the PPC? ABI where function addresses
> aren't (always) the entry point.
> IIRC is causes all sorts of obscure grief.
> 
> I'd also like to know how indirect calls are actually expected to work.
> The whole idea is that the potential targets might be in a kernel module
> that is loaded at run time.
> 
> Scanning a list of possible targets has to be a bad design decision.
> 
> If you are trying to check that the called function has the right
> prototype, stashing a hash of the prototype (or a known random number)
> before the entry point would give most of the benefits without most
> of the costs.
> The linker could be taught to do the same test (much like name mangling
> but without the crap user experience).
> 
> That scheme only has the downside of a data cache miss and (hopefully)
> statically predicted correct branch (hmm - except you don't want to
> speculatively execute the wrong function) and polluting the data cache
> with code.

I admit I was quite surprised by the actual CFI implementation. I would have expected a CFI’d function pointer to actually point to a little descriptor that contains a hash of the function’s type.  The whole bit vector thing seems quite inefficient.

> 
> This all seems like a ploy to force people to buy faster cpus.
> 
>    David
> 
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
>