[openib-general] [PATCH 0/4] Dispatch communication related events to the IB CM

[Rimmer, Todd]

> From: Sean Hefty
> 
> Michael S. Tsirkin wrote:
> > Neither am I, but still I'm a bit uncomfortable - after all,
> > the spec explicitly says:
> >
> > "it should be pointed out that successful reception of the final
message
> > in a CM protocol message exchange cannot be guaranteed."
> 
> The patch handles this by routing the communication establish event
> directly to
> the IB CM for processing, which will force the connection in the
> established
> state and generate a connection event to the user.
> 
> The requirement for RDMA CM users is that they must wait for an
> RDMA_CM_EVENT_ESTABLISHED event before posting sends.  Are we happy
with
> this
> requirement?
> 
> The effect on the user is that either the receive processing must be
> deferred or
> sends must be queued until such an event occurs.  The alternative is
to
> add an
> rdma_establish() call that a user may call after polling their first
> receive
> completion.
> 

An rdma_establish call would not solve the issue.  The problem is the CQ
callbacks are not in a premptable context.  The rdma_establish call
would need to move the QP to RTS, which requires a premptable context.

Todd Rimmer