[ofa-general] detecting "recv buffer not available" events

[Or Gerlitz]

Talpey, Thomas wrote:
> The NFS/RDMA client sets this to 0 when invoking rdma_connect(),
> yet no failures are seen on IB sends even when the server fails to
> promptly post receives.
>   
Just to really make sure you client QP RNR retry counter being set to 
zero, you can issue a QP --query-- after the connection is established 
and see what value is actually configured into it.

Its unclear what you mean by "fails to promptly post receives", if you 
don't post any receive I assume you do get a completion with failure on 
the client send, correct? what's the failure code (wc.status)? AFAIK, 
you must get failures if there is no RX WQE to serve the incoming packet 
and the RNR retries at the sender are set to zero.
> So, does the setting mean anything on the receiving end? I.e., no matter
> what the NFS/RDMA server sets it to, shouldn't we expect the client's
> sends to fail if they don't land on the first try in a receive buffer there?
> That's why the client zeroes rnr_retry - the RPC/RDMA protocol manages
> this so we don't want the transport "helping".
>
>   
Also, from your description it sounds like you need also the server to 
set this to zero when calling rdma_accept

Or