[ofa-general] [PATCH RFC] RDMA/cxgb3: Fixes for zero-stag.

[Steve Wise]

Roland Dreier wrote:
>  > > How about enhancing the iWARP HW drivers implementation of
>  > > get_dma_mr to switch on the access flags provided by the caller, and
>  > > if no remote access it needed, just return zero stag for mr.lkey
>  > > without doing any registration.
>
>  > I think that sounds reasonable.  I can add that for chelsio.
>
> that makes some sense although I don't know of any code that uses the
> zero STag or ib_get_dma_mr() with only local permissions to be honest,
> with the exception of IPoIB, which of course has no relevance to iwarp.
>
>   

NFSRDMA was planning on using the zero stag for recvs on iwarp 
transports.  This boils down to only needing the zero stag + fastreg mrs 
for NFSRDMA server over T3, avoiding dma_mrs which have limits in the T3 HW.

> Also using the zero STag breaks the semantics slightly in that there is
> no longer any PD attached to the key returned by ib_get_dma_mr().
>
>   

True.  I didn't think about this.
> Another point to note is that IB has the analogous concept of "reserved
> L_Key" which we could expose if anyone actually cared.
>   

Hmm.  A new API then?  ib_get_local_lkey() or something?  If IB mandates 
the lkey be value of 0, then we need don't need to do anything...


Steve.