On Fri, Aug 22, 2025 at 08:29:16PM +0000, Qing Zhao wrote:
> > On Aug 22, 2025, at 15:02, Kees Cook <k...@kernel.org> wrote:
> > Right, and sometimes we have to explicitly perform a no-op
> > address-taking to make sure a symbol gets generated:
> >
> > /*
> > * Force the compiler to emit 'sym' as a symbol, so that we can reference
> > * it from inline assembler. Necessary in case 'sym' could be inlined
> > * otherwise, or eliminated entirely due to lack of references that are
> > * visible to the compiler.
> > */
> > #define ___ADDRESSABLE(sym, __attrs)
> > \
> > static void * __used __attrs
> > \
> > __UNIQUE_ID(__PASTE(__addressable_,sym)) = (void *)(uintptr_t)&sym;
> >
> > #define __ADDRESSABLE(sym) \
> > ___ADDRESSABLE(sym, __section(".discard.addressable"))
> >
> > $ git grep KCFI_REFERENCE
> > include/linux/compiler.h:#define KCFI_REFERENCE(sym) __ADDRESSABLE(sym)
> > arch/x86/include/asm/page_64.h:KCFI_REFERENCE(copy_page);
> > arch/x86/include/asm/string_64.h:KCFI_REFERENCE(__memset);
> > arch/x86/include/asm/string_64.h:KCFI_REFERENCE(__memmove);
> > arch/x86/kernel/alternative.c:KCFI_REFERENCE(__bpf_prog_runX);
> > arch/x86/kernel/alternative.c:KCFI_REFERENCE(__bpf_callback_fn);
>
> I am curious on why the compiler eliminates an external routine completely in
> the file if it's address-taken in that file.
> Why an additional no-op address-taken is needed here.
If I am remembering correctly this is needed for rare cases where
a function built without a C definition is being used in Linux's
self-patching "alternatives" code swaps in one function for another,
and is being used indirectly. These cases end up not being visible to
compiler (so no address-taken), but the indirect call site is still
being instrumented. And the above list is the _entire_ list of such
corner cases: all really low-level things.
Peter may remember this better than me...
> > Adding the IPA pass to find all functions worked perfectly. I was able
> > to remove all the weird DECL reconstruction and just use the original
> > FUNCTION_TYPE info for the typeids.
>
> Nice. So this new IPA pass is invoked even when O0?
It seems like it works at any optimization level, yes.
--
Kees Cook
