[openib-general] fmr question

[Caitlin Bestler]

Roland Dreier wrote:
>     Steve> Something similar to the mw_bind semantics should work to
>     Steve> make it more like the iwarp fast-register (i'm not sure
>     Steve> about IB 1.2).  A function like ib_bind_mw() to post the
>     Steve> map WR, and then a new completion type to post the results
> Steve> back to the CQ... 
> 
>     Steve> Would we just change ib_map_phys_fmr() to do this or create
>     Steve> a new API function to preserve backwards compatibility?
> 
> There are pretty big problems with trying to simulate the
> "register memory through a work queue" operation on current
> Mellanox HW (hence the current FMR hack).  As hardware that
> supports the work queue stuff becomes available, I think we
> should just add new APIs (mostly just new work request
> structures) and leave the old FMR interface for Mellanox HW.
> 
>  - R.

That strikes me as the correct approach. We will eventually have
to add device attributes so that ULPs will know which (if any)
of the FMR approaches are supported. And I agree that adding
new work requests/completions is all that will be required
at that later step (beyond the informational attributes).