> > >To: Jason Gunthorpe <jgg@xxxxxxxx> > >From: Steve Wise <swise@xxxxxxxxxxxxxxxxxxxxx> > >Date: 07/16/2018 09:08PM > >Cc: "'Bernard Metzler'" <BMT@xxxxxxxxxxxxxx>, "'Leon Romanovsky'" > ><leon@xxxxxxxxxx>, linux-rdma@xxxxxxxxxxxxxxx > >Subject: Re: [PATCH v3 00/13] Request for Comments on SoftiWarp > > > > > >On 3/1/2018 11:18 AM, Jason Gunthorpe wrote: > >> On Thu, Mar 01, 2018 at 10:28:00AM -0600, Steve Wise wrote: > >>>> On Thu, Mar 01, 2018 at 09:29:04AM -0600, Steve Wise wrote: > >>>> > >>>>> Maybe to add a network interface to a soft-rdma device like rxe, > >we > >>> could > >>>> create a syntax like this: > >>>>> rdma link set rxe_eth0 dev eth0 > >>>> More like: > >>>> > >>>> rdma link add rxe_eth0 type rxe dev eth0 > >>>> > >>>> 'type rxe' triggers the kernel to dispatch to the rxe or siw > >driver to > >>>> create the interface. > >>> Type doesn't sound right though. in the current rdma synax, it > >would be > >>> 'dev'. Maybe: > >>> > >>> rdma link add rxe_eth0 dev rxe netdev eth0 > >> Type is the consistent tag with 'ip link add'. > > > >Hey guys, > > > >I'm starting to think about how to implement 'rdma link add/delete' > >for > >rxe. Does it make sense to add new members to the > rdma_nldev_command > >enum in include/uapi/rdma/rdma_netlink.h for adding and deleting > >soft-rdma links to netdev interfaces? Something like > >RDMA_NLDEV_CMD_LINK_ADD and RDMA_NLDEV_CMD_LINK_DEL? Then > add > >handlers > >to the nldev_cb_table array in drivers/infiniband/core/nldev.c. > >Can > >we assume there is only one soft-rdma driver for each rdma > >transport? > >That would enable a simple core->driver dispatch via an array of > >driver-specific handler functions that are indexed by the soft-rdma > >transport type. Or do we need something more general? > > > >Just thinking out loud here. What do you all think? > > > Hey Steve, > > Would be great if you would bring that thing forward. Highly > appreciated! > I would as well assume only one software rdma driver per > rdma transport. But in your proposed syntax, there is a > driver name (rxe) anyway, so multiple drivers could also be > selectable...? Actually, shall we specify the rdma transport at > all, or is that a property the device is providing during > registration? I would keep it simple for now and just bind soft rdma driver X with netdev Y and let the driver's defaults take hold. Once we iron that out and have some code to look at, we will probably have more ideas on how to tune the design.. Steve. -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
