Re: [PATCH] x86: modify inline asm constraints in __cmpxchg_double()

[Juergen Gross <jgross@suse.com>]

On 16/10/2018 10:01, Ingo Molnar wrote:
> 
> * Juergen Gross <jgross@suse.com> wrote:
> 
>> On 16/10/2018 08:25, Ingo Molnar wrote:
>>>
>>> * Juergen Gross <jgross@suse.com> wrote:
>>>
>>>> Some gcc versions seem to have problems with the constraints in
>>>> __cmpxchg_double() as they suddenly issue a build error when random
>>>> parts of sources calling __cmpxchg_double() are modified, like e.g.
>>>> slub.c. This has been observed on Debian systems only so far.
>>>>
>>>> Using "0" instead of "a" in the input constraints has the same
>>>> semantics while avoiding that build error.
>>>>
>>>> Signed-off-by: Juergen Gross <jgross@suse.com>
>>>> ---
>>>> I should note that I have observed gcc hangs instead sometimes. Not
>>>> taking any patches modifying users of __cmpxchg_double() due to a gcc
>>>> bug which seems to be distro-specific is a bad move IMO.
>>>>
>>>> I'd rather make it clear from build behavior that this is a bug in gcc
>>>> by letting the build hang instead of throwing error warnings not in any
>>>> way related to changes in the code.
>>>>
>>>> The gcc bug is filed under:
>>>>
>>>> https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=908154
>>>> ---
>>>>  arch/x86/include/asm/cmpxchg.h | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/arch/x86/include/asm/cmpxchg.h b/arch/x86/include/asm/cmpxchg.h
>>>> index a55d79b233d3..b3b4d61a8969 100644
>>>> --- a/arch/x86/include/asm/cmpxchg.h
>>>> +++ b/arch/x86/include/asm/cmpxchg.h
>>>> @@ -245,7 +245,7 @@ extern void __add_wrong_size(void)
>>>>  	asm volatile(pfx "cmpxchg%c4b %2; sete %0"			\
>>>>  		     : "=a" (__ret), "+d" (__old2),			\
>>>>  		       "+m" (*(p1)), "+m" (*(p2))			\
>>>> -		     : "i" (2 * sizeof(long)), "a" (__old1),		\
>>>> +		     : "i" (2 * sizeof(long)), "0" (__old1),		\
>>>>  		       "b" (__new1), "c" (__new2));			\
>>>
>>> This got changed to +a in -tip:
>>>
>>>   c808c09b527c: x86/asm: Use CC_SET()/CC_OUT() in __cmpxchg_double()
>>>
>>> Mind sending a patch on top of -tip?
>>
>> No, I think you can drop my patch.
>>
>> c808c09b527c: x86/asm: Use CC_SET()/CC_OUT() in __cmpxchg_double()
>> has the same effect as my patch: the build failure is replaced by
>> gcc hanging instead in the critical configuration.
> 
> Ok. No sane workaround we can do to avoid the hang I suspect, right?

I'm not aware of any.

> Should we propagate the fix/enhancement to x86/urgent? It's currently in 
> x86/asm for a v4.20 merge.

My patch triggering the issue is queued for 4.20, too.

There might be other constellations triggering it, but I haven't
seen any yet.


Juergen