Re: [PATCH RFC 2/2] RDMA/nldev: provide detailed CM_ID information

[Leon Romanovsky <leon-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 1562 bytes --]

On Mon, Feb 05, 2018 at 09:33:07AM -0600, Steve Wise wrote:
> Hey Leon,
>
> >
> > It is opposite, cm_id is a unique number, but other objects don't have
> > such. What about PD and CQ?
> >
> > We can declare that access to QP .doit willbe based on QPN, PD .doit
> > will be based on local_dma_lkey, but what will be CQ identifier?
> >
>
> Now that I've thought about this more, I realize the core verbs data
> structures don't have identifiers for most objects.  I'm not sure why the
> qp_num is exposed, but I guess it is because the qp num values must be
> exchanged between both ends of an IB connection to setup the connection.
>
> The  local_dma_lkey  value will not be unique for each PD.  For all iwarp
> devices the local_dma_lkey is 0.  (I think the same is true for IB devices).
> So PD has the same issue as CQ.
>
> I suppose we can use rkey for MRs, but if the MR is not in the VALID state,
> the rkey isn't necessarily accurate (the LSB of the rkey can be changed by
> the application with each REG_MR operation).
>
> The cm_id doesn't have any unique identifier either, except maybe the
> combination of src/dst addresses and port space.
>
> With iw_cxgb4, CQs, PDs, and MRs all have a unique identifier just like a QP
> does.  I expect all hw implementations of rdma have numeric identifiers for
> CQs, PDs, QPs, etc.    They aren't exposed to the core API though.
>
> Perhaps we add a device-unique (or globally unique) identifier as part of
> the restrack struct and use that for GET/SET?

+1,
I like it, it is very clean.

Thanks

>
> Steve.
>

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]