Re: [RFC PATCH net 1/2] [BUGFIX] bonding: use local function pointer of bond->recv_probe in bond_handle_frame

[HAYASAKA Mitsuo <mitsuo.hayasaka.hu@hitachi.com>]

Hi Eric,

Thank you for your comment.

(2011/10/11 22:23), Eric Dumazet wrote:
> Le mardi 11 octobre 2011 à 22:02 +0900, HAYASAKA Mitsuo a écrit :
>> Hi WANG Cong
>>
>> Thank you for your comments.
>>
>> (2011/10/07 22:24), Américo Wang wrote:
>>> On Fri, Oct 7, 2011 at 8:49 PM, Mitsuo Hayasaka
>>> <mitsuo.hayasaka.hu@hitachi.com> wrote:
>>>> The bond->recv_probe is called in bond_handle_frame() when
>>>> a packet is received, but bond_close() sets it to NULL. So,
>>>> a panic occurs when both functions work in parallel.
>>>>
>>>> Why this happen:
>>>> After null pointer check of bond->recv_probe, an sk_buff is
>>>> duplicated and bond->recv_probe is called in bond_handle_frame.
>>>> So, a panic occurs when bond_close() is called between the
>>>> check and call of bond->recv_probe.
>>>>
>>>> Patch:
>>>> This patch uses a local function pointer of bond->recv_probe
>>>> in bond_handle_frame(). So, it can avoid the null pointer
>>>> dereference.
>>>>
>>>
>>> Hmm, I don't doubt it can fix the problem, I am wondering if
>>> bond->recv_probe should be protected by bond->lock...
>>
>> Indeed, in general any resources should be protected from the asynchronous
>> workers.
>>
>> At first, I thought it should be handled with lock protection, as well.
>> However, I guess that using bond->lock on this kind of hot-path may
>> introduces unnecessary overhead. In addition, this code works well
>> without the strict lock protection. So, I think this change is the
>> right way to fix it.
> 
> Maybe, but then ACCESS_ONCE() is needed to prevent compiler to
> 'optimize' the temporary variable.
> 
> Or use rcu_dereference() to make the whole thing really safe and self
> documented.
> 
I agreed.
I'd like to send the patch with ACCESS_ONCE(), again.

Thanks.