[openib-general] Re: CMA backlog

[Fabian Tillier]

On 5/30/06, Michael S. Tsirkin <mst at mellanox.co.il> wrote:
> Quoting r. Sean Hefty <sean.hefty at intel.com>:
> > >Currently, this is not something that can be implemented by ULP on top of
> > >CMA, because returning error from REQ will result in reject rather than REQ
> > >drop.
> >
> > A generic ULP could handle this by making use of the private data, and retrying
> > requests after a REJ with insufficient resources.
>
> You are right in this, of course. What I meant is that dropping REQ
> can not be implemented by ULPs without extending our CMA and CM.
>
> As you point out, retrying on reject might make sense for some
> ULPs, but happily these can already implement this without extending CMA
> and CM. So we are set there.
>
> I suggest adding an option of dropping REQ, emulating TCP behaviour.
> This is IMO required for SDP, as I read the spec and from I know about other
> SDP implementations.

You mean a half-reject, where it only rejects locally but doesn't send
the REJ?  This shouldn't be that hard to do if the local CM supported
a special reject code that would suppress the REJ MAD being sent.

Alternatively, at least in Windows, the CM will continue to retry REQ
requests if it receives a REJ with INVALID_SID as the reason (i.e. the
app is not up yet).  If the Linux CM could do the same, we could use
this reject reason to "drop" the REQ.

- Fab