[Openib-windows] Fence after RDMA read

[Fabian Tillier]

Hi Tzachi,

On 6/18/06, Tzachi Dar <tzachid at mellanox.co.il> wrote:
>
> Hi Fab,
>
> While debugging a failure in our regression system, we have found that the
> current code has a fence in the next send after RDMA read. There is no
> arguing that currently this fence is needed.
>
> Still I wander if this is something that is documented, is this a workaround
> for a bug in switch, or are we hiding some other bug here?

The docs aren't all that clear - the switch will issue the RDMA read,
and immediately post the read complete message to the send queue.  The
provider doesn't have a vway to tell which send is the message
complete indication, so the first send after the RDMA read is always
fenced.

- Fab