Re: [PATCH 1/1] alloc_tag: work around clang-14 issue with __builtin_object_size()

[Nathan Chancellor]

On Wed, Feb 05, 2025 at 11:18:35AM -0800, Kees Cook wrote:
> On Sat, Feb 01, 2025 at 12:05:03PM -0800, Suren Baghdasaryan wrote:
> > Additional condition in the allocation hooks causes Clang version 14
> > (tested on 14.0.6) to treat the allocated object size as unknown at
> > compile-time (__builtin_object_size(obj, 1) returns -1) even though
> > both branches of that condition yield the same result. Other versions
> > of Clang (tested with 13.0.1, 15.0.7, 16.0.6 and 17.0.6) compile the
> > same code without issues. Add build-time Clang version check which
> > removes this condition and effectively restores the unconditional tag
> > store/restore flow when compiled with clang-14.
> > 
> > Fixes: 07438779313c ("alloc_tag: avoid current->alloc_tag manipulations when profiling is disabled")
> > Reported-by: kernel test robot <lkp@xxxxxxxxx>
> > Closes: https://lore.kernel.org/oe-kbuild-all/202501310832.kiAeOt2z-lkp@xxxxxxxxx/
> > Signed-off-by: Suren Baghdasaryan <surenb@xxxxxxxxxx>
> > ---
> >  include/linux/alloc_tag.h | 15 ++++++++++++++-
> >  1 file changed, 14 insertions(+), 1 deletion(-)
> > 
> > diff --git a/include/linux/alloc_tag.h b/include/linux/alloc_tag.h
> > index a946e0203e6d..df432c2c3483 100644
> > --- a/include/linux/alloc_tag.h
> > +++ b/include/linux/alloc_tag.h
> > @@ -222,10 +222,23 @@ static inline void alloc_tag_sub(union codetag_ref *ref, size_t bytes) {}
> >  
> >  #endif /* CONFIG_MEM_ALLOC_PROFILING */
> >  
> > +/* See https://lore.kernel.org/all/202501310832.kiAeOt2z-lkp@xxxxxxxxx/ */
> > +#if defined(CONFIG_CC_IS_CLANG) && CONFIG_CLANG_VERSION >= 140000 && CONFIG_CLANG_VERSION < 150000
> 
> FWIW, this could just be "< 150000" -- < 14 doesn't warn because (as
> Nathan mentioned to me today) it didn't support the build-time error
> attribute, so it wouldn't have warned even if it did trip over it.
> 
> > +static inline bool store_current_tag(void)
> > +{
> > +	return true;
> > +}
> > +#else
> > +static inline bool store_current_tag(void)
> > +{
> > +	return mem_alloc_profiling_enabled();
> > +}
> > +#endif
> > +
> >  #define alloc_hooks_tag(_tag, _do_alloc)				\
> >  ({									\
> >  	typeof(_do_alloc) _res;						\
> > -	if (mem_alloc_profiling_enabled()) {				\
> > +	if (store_current_tag()) {					\
> >  		struct alloc_tag * __maybe_unused _old;			\
> >  		_old = alloc_tag_save(_tag);				\
> >  		_res = _do_alloc;					\
> 
> I think the work-around is fine, but I'm trying to dig into the root
> cause here.
> 
> As you found, it fails on the final strtomem_pad:
> 
> 	strtomem_pad(key->u.kbd.press_str, press, '\0');
> 	strtomem_pad(key->u.kbd.repeat_str, repeat, '\0');
> 	strtomem_pad(key->u.kbd.release_str, release, '\0');
> 
> (but not the earlier calls??) The destinations are:
> 
> 		char press_str[sizeof(void *) + sizeof(int)] __nonstring;
> 		char repeat_str[sizeof(void *) + sizeof(int)] __nonstring;
> 		char release_str[sizeof(void *) + sizeof(int)] __nonstring;
> 
> Random thoughts include "this is the last array in the struct" which might
> imply bad compiler behavior about its sizing via __builtin_object_size()
> (i.e. trailing array must always be unknown size to deal with
> fake flex arrays), but that wasn't fixed until Clang 16 (with
> -fstrict-flex-arrays=3), so that it doesn't trip in Clang 15 is odd.

I bisected the fix in LLVM 15 to
https://github.com/llvm/llvm-project/commit/d8e0a6d5e9dd2311641f9a8a5d2bf90829951ddc,
which certainly makes sense. Given the commit mentions phi nodes and
folding means that maybe there is a branch that was not getting
eliminated before this change? I have not really looked into the call
chain here.

> To Kent's comment[1], I believe I was using __builtin_object_size() here
> because I have a knee-jerk aversion to sizeof() due to it blowing up on
> flexible arrays, but that's not relevant here. ARRAY_SIZE() would work,
> but only if type checking to "char *" succeeds, as Kent suggests.
> 
> Let me see if making those changes survives testing...

If that suggestion works, I would certainly prefer that to a compiler
version workaround. Worst case, we could bump the minimum supported LLVM
version over this but it does not seem serious enough to do so at the
moment.

Cheers,
Nathan