[openib-general] [PATCH] kDAPL: cleanup dat/ a bit more

[Caitlin Bestler]

On 6/3/05, Christoph Hellwig <hch at lst.de> wrote:

> 
> > Keep in mind that loading/unloading DAT Provider is *not*
> > synonymous with loading/unloading drivers. In fact I believe
> > the intent is to have a single provider that supports multiple
> > devices. Such a provider would simply register itself when
> > it was loaded.
> 
> The current code looks like that, yes.  But long term there should
> be no separate DAT provider, but the functionality now covered by
> kDAP should be converged into a general RDMA provider.
> 

This issue was discussed in OpenRDMA, and reached the same
goal. Yes, in the long run there is no need for there to be both
dynamic loading of DAT Providers and of Verb Providers. But
that is a necessary interim transition.

OpenRDMA discussed with the DAT Collaborative the idea
of subsuming the responsibilities of the DAT Registry, so that
the OpenRDMA directory could take the 'dat_xxx' calls
directly. When the device dependent logic used a dynamically
loaded verbs, those would be invoked directly. When the
device dependent logic required a full DAT Provider (because
it did not provide a recognized verbs interface) then the
existing dynamic DAT dispatch would be used.

That approach is certainly applicable for OpenIB as well.
The key is recognizing the need for a transition plan.
Customers have DAT Providers installed now, they
cannot synchronize getting new DAT Providers from
their suppliers with a new Linux release. This is
especially true since OpenIB does not currently
define a verbs interface that is suitable for iWARP
vendors to use. So dat_ia_openv() needs to still
support existing dat_registry logic and existing
DAT_PROVIDERs, otherwise it is breaking 
existing code.