[ofa-general] Re: IB/IPoIB Check if grat. AR changed had arrived when working in connected mode

[Eli Cohen]

On Thu, 2008-01-31 at 08:48 +0200, Or Gerlitz wrote:
> Eli Cohen wrote:
> >> Eli Cohen wrote:
> >>> Now you may call ipoib_put_ah(neigh->ah) for a CM neighbor and this
> >>> could cause de-reference of a NULL pointer.
> 
> Eli,
> 
> Not that your assumption changes anything regarding the patch 
> correctness (and if it does, please suggest what problem is introduced 
> by the patch), but, you might have some misconception here that for a 
> connected mode neighbour neigh->ah is NULL.
> 
> This is wrong, see patch_rec_completion where first an address handle is 
> created and then assigned to the neighbour no matter if its a connected 
> mode or datagram mode one.
Yes, thanks for pointing this out.

> 
> One might say that there's a resource waste here, since for connected 
> mode neighbour the driver consumes two HCA resources (TX QP and AH) 
> where for datagram mode neigh it consumes only one. This seems as not 
> too easy cleanup, which maybe best left for the future generations...
> 

I think the waste of resources is not that significant and we have the
benefit that it allows switching smoothly from CM to UD mode.