[openib-general] [PATCH] sa_query: require SA query registration

[Sean Hefty]

>> * Move struct ib_sa_client definition internal to SA module to
>>   better encapsulate future extensions.
>
>Documentation/stable_api_nonsense.txt says:
>"You think you want a stable kernel interface, but you really do not"

Encapsulation is typically a good idea.  The user cannot modify anything inside
of struct ib_sa_client without breaking it.  The only benefit we gain by
exposing it is saving a memory allocation.

The verb interfaces expose the structures to allow for direct calls from a ULP
to a provider for speed path operations.

>Why is
>	ipoib_sa_client = ib_sa_register_client
>	ret = PTR_ERR(ipoib_sa_client)
>a better encapsulation than
>	ret = ib_sa_register_client(&ipoib_sa_client)

The former completely hides the internal definition of the structure from the
user.  The latter does not, providing no encapsulation.

>Further, why should this API be different from ib_register_client?
>Isn't this confusing?

It's different from ib_register_client, but more similar to
ib_register_mad_agent.  If anything it would make more sense to be closer to
ib_register_mad_agent.  For example, a callback function makes more sense passed
into a register call than every query call.

>Isn't it better to let user control where does the memory come from,
>rather than forcing GFP_KERNEL?

The sa_query module _is_ the user of the memory, but can add a gfp_mask if we
think that's useful.

- Sean