Re: Re: Invalidation

> No.  It appears that invalidation is a rather important mechanism
> for actually using STags, and so, aside from historical reasons, what's
> the rationale for it being an RDMAP operation rather than a DDP
> operation?

Caitlin said it: DDP is only about placement and RDMA is about
control.  [But, by the time we're done who knows anymore.]

Invalidation is an important mechanism for CERTAIN applications.  And
given that all RDMA-family protocols over the past 10 or so have
pretty much done it the same way [cough, gasp, not...], I'm sure we
know exactly how to do it.  We should commit it irreversably to gates
and provide no alternatives.

DDP should be a MINIMUM, self-contained placement mechanism.  It is a
base upon which different RDMA-type protocols can be built.  ULPs can
also use DDP, but they should not expect to get a `RDMA-lite'
semantics.  RDMA is whatever fuller placement-based protocol the
community wants for current & anticipated ULPs.  As the layering line
on DDP is pushed down as far as possible, RDMA's service interface
picks up the slack.  For example, DDP should not be saying anything
about how host-side untagged message resources are managed
(i.e. ddp_post_recv()).  But it does.  Oh well.

You can build a variety of invalidation mechanisms on top of DDP.
These could be built by a DDP-targeted ULP, or by an RDMAP (a `general
purpose' DDP ULP).  Invalidation mechanisms defined by an RDMAP become
a candidates for hardware handling (performance optimization) like
write with invalidate.

Steph