Re: [PATCH v6 1/4] Compiler Attributes: Add __pass_object_size for Clang

From: Kees Cook
Date: Thu Feb 03 2022 - 15:58:22 EST


On Thu, Feb 03, 2022 at 12:18:24PM -0800, Nick Desaulniers wrote:
> On Thu, Feb 3, 2022 at 9:33 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > In order to gain greater visibility to type information when using
> > __builtin_object_size(), Clang has a function attribute "pass_object_size"
> > that will make size information available for marked arguments in
> > a function by way of implicit additional function arguments that are
> > then wired up the __builtin_object_size().
> >
> > This is needed to implement FORTIFY_SOURCE in Clang, as a workaround
> > to Clang's __builtin_object_size() having limited visibility[1] into types
> > across function calls (even inlines).
> >
> > Since any usage must also be const, include it in the macro.
>
> I really don't like hiding the qualifier in the argument-attribute
> macro like that. I'd rather we be more explicit about changing the
> function interfaces in include/linux/fortify-string.h.

It seemed redundant to have it separate when it would always be needed.
In v5 I had the const in the BOS (now POS) macro, but realized that this
was silly since it would always need to be that way for __pos.

> For instance, in patch 4/4, let's take a look at this hunk:
>
> -__FORTIFY_INLINE char *strncpy(char *p, const char *q, __kernel_size_t size)
> ...
> +__FORTIFY_INLINE __diagnose_as(__builtin_strncpy, 1, 2, 3)
> +char *strncpy(char * POS p, const char *q, __kernel_size_t size)
>
> manually expanded, this has changed the qualifiers on the type of the
> first parameter from `char *` to `char * const`.

Yup, that's expected, and I wanted to tie it strictly to the expansion
of __pass_object_size since otherwise GCC would gain the "const" bit
(which technically shouldn't matter, but this whole series has been
nothing but corner case after corner case, and I didn't want to risk
another one).

> I think it might be helpful to quote the docs
> (https://clang.llvm.org/docs/AttributeReference.html#pass-object-size-pass-dynamic-object-size)
>
> >> Additionally, any parameter that pass_object_size is applied to must be marked const at its function’s definition.
>
> One thing that's concerning to me is though:
>
> >> It is an error to take the address of a function with pass_object_size on any of its parameters.
>
> Surely the kernel has indirect calls to some of these functions
> somewhere? Is that just an issue for C++ name-mangling perhaps?

AFAIU, this shouldn't be a problem for any of these. Nothing is trying
to take memcpy, memset, etc by address. The macro-ified version of this
change proved that out. :)

--
Kees Cook