Re: [PATCH] net: optimize cmpxchg in ip_idents_reserve

[Shaokun Zhang]

Hi Peter/Eric,

Shall we use atomic_add_return() unconditionally and add some comments? Or I missed
something.

Thanks,
Shaokun

On 2020/1/20 16:18, Peter Zijlstra wrote:
> On Fri, Jan 17, 2020 at 10:48:19AM -0800, Eric Dumazet wrote:
>>
>>
>> On 1/17/20 10:38 AM, Arvind Sankar wrote:
>>> On Fri, Jan 17, 2020 at 10:16:45AM -0800, Eric Dumazet wrote:
>>>> WasÅt it the case back in 2016 already for linux-4.8 ?
>>>>
>>>> What will prevent someone to send another report to netdev/lkml ?
>>>>
>>>>  -fno-strict-overflow support is not a prereq for CONFIG_UBSAN.
>>>>
>>>> Fact that we kept in lib/ubsan.c and lib/test_ubsan.c code for 
>>>> test_ubsan_add_overflow() and test_ubsan_sub_overflow() is disturbing.
>>>>
>>>
>>> No, it was bumped in 2018 in commit cafa0010cd51 ("Raise the minimum
>>> required gcc version to 4.6"). That raised it from 3.2 -> 4.6.
>>>
>>
>> This seems good to me, for gcc at least.
>>
>> Maybe it is time to enfore -fno-strict-overflow in KBUILD_CFLAGS 
>> instead of making it conditional.
> 
> IIRC there was a bug in UBSAN vs -fwrapv/-fno-strict-overflow that was
> only fixed in gcc-8 or 9 or so.
> 
> So while the -fwrapv/-fno-strict-overflow flag has been correctly
> supported since like forever, UBSAN was buggy until quite recent when
> used in conjustion with that flag.
> 
> .
>