> From: Jason Gunthorpe [mailto:[email protected]]
> Sent: Thursday, June 11, 2015 2:52 PM
> To: Wan, Kaike
> Cc: [email protected]; Fleck, John; Weiny, Ira
> Subject: Re: [PATCH v5 1/4] IB/netlink: Add defines for local service requests
> through netlink
>
> On Thu, Jun 11, 2015 at 06:32:44PM +0000, Wan, Kaike wrote:
> > > From: Jason Gunthorpe [mailto:[email protected]]
> > > Sent: Thursday, June 11, 2015 1:58 PM
> > > To: Wan, Kaike
> > > Cc: [email protected]; Fleck, John; Weiny, Ira
> > > Subject: Re: [PATCH v5 1/4] IB/netlink: Add defines for local
> > > service requests through netlink
> > >
> > > On Thu, Jun 11, 2015 at 01:03:42PM -0400, [email protected] wrote:
> > > > From: Kaike Wan <[email protected]>
> > > >
> > > > This patch adds netlink defines for SA client, local service
> > > > group, local service operations, and related attributes.
>
> > + LS_NLA_TYPE_REVERSIBLE,
> > + LS_NLA_TYPE_NUMB_PATH,
>
> These need to be combined. It does not make sense to request numb_paths
> when the API we have is designed to return APM data.
>
> Recommend a 'QP Type' with options of:
> - RC QP: Return up to the full 6 path tuple with full APM data
> - UD QP: Return a single non-reversible path
> - GMP QP: Return up to two reversible GMP paths.
>
> If it is hard to get accurate information then you can sketch this in and
> infer
> RC QP and GMP QP by the request having the reversible set bit. But ideally
> RDMA CM and SRP would request RC QP, IPoIB would request UD QP.
It's quite indirect here for this implementation. (ib_sa_path_rec_get())
Caller reversible numb_path QP Type
--------------------------------------------------------------------------
cma 1 1
GMP
ipoib Not set 1 (?)
UD
srp Not set 1
UD
In addition, on the user side, we need to convert QP type back into
reversible_numpath again, which may not interpret the original request
correctly.
Why should we do all the translation?
Is QP type definition already exported to user space?
I see IBV_QPT_RC/UC/UD/RAW_PACKET/XRC_SEND/XRC_RECV, but not GMP in verbs.h.
>
> > +/* Local Service ServiceID attribute */ struct rdma_nla_ls_service_id
> > +{
> > + __be64 service_id;
> > +};
>
> Do not expose BE to userspace, everything should be in cpu order.
If we use cpu order, we need to do two conversions: from BE to cpu order in
kernel and from cpu order to BE in user space. Struct ib_user_path_rec contains
many __be32 fields.
>
> Jason
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
