Re: [PATCH v4 07/14] arm64: Add support for suppressing warning backtraces

Le 18/03/2025 à 16:59, Will Deacon a écrit :
> On Thu, Mar 13, 2025 at 05:40:59PM +0100, Alessandro Carminati wrote:
>> On Thu, Mar 13, 2025 at 1:25 PM Will Deacon <[email protected]> wrote:
>>>
>>> On Thu, Mar 13, 2025 at 11:43:22AM +0000, Alessandro Carminati wrote:
>>>> diff --git a/arch/arm64/include/asm/bug.h b/arch/arm64/include/asm/bug.h
>>>> index 28be048db3f6..044c5e24a17d 100644
>>>> --- a/arch/arm64/include/asm/bug.h
>>>> +++ b/arch/arm64/include/asm/bug.h
>>>> @@ -11,8 +11,14 @@
>>>>
>>>>   #include <asm/asm-bug.h>
>>>>
>>>> +#ifdef HAVE_BUG_FUNCTION
>>>> +# define __BUG_FUNC  __func__
>>>> +#else
>>>> +# define __BUG_FUNC  NULL
>>>> +#endif
>>>> +
>>>>   #define __BUG_FLAGS(flags)                           \
>>>> -     asm volatile (__stringify(ASM_BUG_FLAGS(flags)));
>>>> +     asm volatile (__stringify(ASM_BUG_FLAGS(flags, %c0)) : : "i" (__BUG_FUNC));
>>>
>>> Why is 'i' the right asm constraint to use here? It seems a bit odd to
>>> use that for a pointer.
>>
>> I received this code as legacy from a previous version.
>> In my review, I considered the case when HAVE_BUG_FUNCTION is defined:
>> Here, __BUG_FUNC is defined as __func__, which is the name of the
>> current function as a string literal.
>> Using the constraint "i" seems appropriate to me in this case.
>>
>> However, when HAVE_BUG_FUNCTION is not defined:
>> __BUG_FUNC is defined as NULL. Initially, I considered it literal 0,
>> but after investigating your concern, I found:
>>
>> ```
>> $ echo -E "#include <stdio.h>\n#include <stddef.h>\nint main()
>> {\nreturn 0;\n}" | aarch64-linux-gnu-gcc -E -dM - | grep NULL
>> #define NULL ((void *)0)
>> ```
>>
>> I realized that NULL is actually a pointer that is not a link time
>> symbol, and using the "i" constraint with NULL may result in undefined
>> behavior.
>>
>> Would the following alternative definition for __BUG_FUNC be more convincing?
>>
>> ```
>> #ifdef HAVE_BUG_FUNCTION
>>      #define __BUG_FUNC __func__
>> #else
>>      #define __BUG_FUNC (uintptr_t)0
>> #endif
>> ```
>> Let me know your thoughts.
> 
> Thanks for the analysis; I hadn't noticed this specific issue, it just
> smelled a bit fishy. Anyway, the diff above looks better, thanks.

That propably deserves a comment.

Doesn't sparse and/or checkpatch complain about 0 being used in lieu of 
NULL ?

By the way I had similar problem in the past with GCC not seeing NULL as 
a __builtin_constant_p(), refer commit 1d8f739b07bd ("powerpc/kuap: Fix 
set direction in allow/prevent_user_access()")

Christophe