[ofa-general] Re: [RFC] [PATCH v3] IB/ipoib: Add bonding support to IPoIB

[Moni Shoua]

>>> Using to_ipoib_neigh outside priv->lock looks problematic.
>>> Can you convince me this does not introduce new races?
>>>
>>>
>> I can try...
>> ipoib_neigh_destructor is called from neigh_destroy() and this is when the
>> kernel neighbour is under destruction itself and no one holds a reference to
>> it. 
> 
> OK but we might have references to ipoib_neigh. Specifically path and mcast
> group all might have it - that's what neigh_list is.
> 
Maybe I'm not get something but how does the presence of ipoib_neigh on the list is a problem?
to_ipoib_neigh() takes the pointer from the neighbour itself without caring if it is on a list or not.
Destruction itself is being done under lock.

>> My opinion is that if I can't assume that no one is touching ipoib_neigh when kernel 
>> neighbour is being destroyed then we have a bigger problem.
> 
> That's what locks you remove seem to be there for.
>