Re: linux-next: Tree for Dec 6 (objtool, lots in btrfs)

[Randy Dunlap <rdunlap@infradead.org>]

On 12/13/19 3:50 PM, Josh Poimboeuf wrote:
> On Fri, Dec 13, 2019 at 03:03:11PM -0800, Randy Dunlap wrote:
>> On 12/12/19 12:25 PM, Randy Dunlap wrote:
>>> On 12/12/19 10:47 AM, Josh Poimboeuf wrote:
>>>> On Wed, Dec 11, 2019 at 08:21:38AM -0800, Randy Dunlap wrote:
>>>>> [oops, forgot to add Josh and PeterZ]
>>>>>
>>>>> On 12/11/19 5:49 AM, David Sterba wrote:
>>>>>> On Fri, Dec 06, 2019 at 08:17:30AM -0800, Randy Dunlap wrote:
>>>>>>> On 12/5/19 6:54 PM, Stephen Rothwell wrote:
>>>>>>>> Hi all,
>>>>>>>>
>>>>>>>> Please do not add any material for v5.6 to your linux-next included
>>>>>>>> trees until after v5.5-rc1 has been released.
>>>>>>>>
>>>>>>>> Changes since 20191204:
>>>>>>>>
>>>>>>>
>>>>>>> on x86_64:
>>>>>>>
>>>>>>> fs/btrfs/ctree.o: warning: objtool: btrfs_search_slot()+0x2d4: unreachable instruction
>>>>>>
>>>>>> Can somebody enlighten me what is one supposed to do to address the
>>>>>> warnings? Function names reported in the list contain our ASSERT macro
>>>>>> that conditionally calls BUG() that I believe is what could cause the
>>>>>> unreachable instructions but I don't see how.
>>>>>>
>>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/fs/btrfs/ctree.h#n3113
>>>>>>
>>>>>> __cold
>>>>>> static inline void assfail(const char *expr, const char *file, int line)
>>>>>> {
>>>>>> 	if (IS_ENABLED(CONFIG_BTRFS_ASSERT)) {
>>>>>> 		pr_err("assertion failed: %s, in %s:%d\n", expr, file, line);
>>>>>> 		BUG();
>>>>>> 	}
>>>>>> }
>>>>>>
>>>>>> #define ASSERT(expr)	\
>>>>>> 	(likely(expr) ? (void)0 : assfail(#expr, __FILE__, __LINE__))
>>>>>>
>>>>
>>>> Randy, can you share one of the btrfs .o files?  I'm not able to
>>>> recreate.
>>>>
>>>
>>> Hm. I'll have to try to recreate this. I no longer have files from 20191206
>>> (lack of space).
>>>
>>> I'll let you know if/when I can recreate it.
>>
>> OK, 40 builds later, I have reproduced it.
>>
>> I am attaching one of the btrfs .o files and the kernel config file (FTR).
>> (gzipped)
>> Let me know if you want more of the .o files.
> 
> Thanks.  This is arguably a compiler bug, but the below produces better
> code generation by adding a noreturn annotation.  I think GCC gets
> tripped up by the IS_ENABLED conditional and can't always tell that
> assfail (sic) doesn't return.
> 
> BTW, I'm on my way out the door for a week of much-needed PTO but I can
> handle this patch (and several others I have pending which were reported
> by you) when I get back.

Sure, no hurry.  Have a good one.

> diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
> index b2e8fd8a8e59..bbd68520f5f1 100644
> --- a/fs/btrfs/ctree.h
> +++ b/fs/btrfs/ctree.h
> @@ -3110,14 +3110,16 @@ do {								\
>  	rcu_read_unlock();					\
>  } while (0)
>  
> -__cold
> +#ifdef CONFIG_BTRFS_ASSERT
> +__cold __unlikely

what provides __unlikely?  It is causing build errors.

and if I remove the "__unlikely", I still see the objtool warnings
(unreachable instructions).

>  static inline void assfail(const char *expr, const char *file, int line)
>  {
> -	if (IS_ENABLED(CONFIG_BTRFS_ASSERT)) {
> -		pr_err("assertion failed: %s, in %s:%d\n", expr, file, line);
> -		BUG();
> -	}
> +	pr_err("assertion failed: %s, in %s:%d\n", expr, file, line);
> +	BUG();
>  }
> +#else
> +static inline void assfail(const char *expr, const char *file, int line) {}
> +#endif
>  
>  #define ASSERT(expr)	\
>  	(likely(expr) ? (void)0 : assfail(#expr, __FILE__, __LINE__))
> 


-- 
~Randy