Re: [PATCH v2 03/13] Attach/detach SoftiWarp to/from network and RDMA subsystem

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

[-- Attachment #1: Type: text/plain, Size: 2945 bytes --]

On Thu, Nov 09, 2017 at 02:43:41PM +0000, Bernard Metzler wrote:
>
> -----Leon Romanovsky <leon-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org> wrote: -----
>
> >To: Bernard Metzler <BMT-OA+xvbQnYDHMbYB6QlFGEg@public.gmane.org>
> >From: Leon Romanovsky <leon-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org>
> >Date: 11/09/2017 09:52AM
> >Cc: linux-rdma-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
> >Subject: Re: [PATCH v2 03/13] Attach/detach SoftiWarp to/from network
> >and RDMA subsystem
> >
> >On Wed, Nov 08, 2017 at 04:46:26PM +0000, Bernard Metzler wrote:
> >> -----Leon Romanovsky <leon-DgEjT+Ai2ygdnm+yROfE0A@public.gmane.org> wrote: -----
> >>
> >> >> + */
> >> >> +static __init int siw_init_module(void)
> >> >> +{
> >> >> +	int rv;
> >> >> +	int nr_cpu;
> >> >> +
> >> >> +	if (SENDPAGE_THRESH < SIW_MAX_INLINE) {
> >> >> +		pr_info("siw: sendpage threshold too small: %u\n",
> >> >> +			(int)SENDPAGE_THRESH);
> >> >> +		rv = EINVAL;
> >> >> +		goto out;
> >> >> +	}
> >> >> +	/*
> >> >> +	 * The xprtrdma module needs at least some rudimentary bus to
> >set
> >> >> +	 * some devices path MTU.
> >> >> +	 */
> >> >> +	rv = bus_register(&siw_bus);
> >> >
> >> >bus register for the driver? no way.
> >>
> >> I admit - this looks ugly. And the comment above is incomplete.
> >> I did this mainly to allow siw to register with loopback devices.
> >> During device registration, we need to present a real parent
> >device:
> >> device_add() needs it, as called by ib_device_register_sysfs().
> >> The loopback device does not have a parent device, nor dma_ops.
> >> The dma_ops can be satisfied by global dma_virt_ops. But I did not
> >> find a similar 'virtual parent device' I could present during
> >registration.
> >> So I 'invented' that one.
> >>
> >> I see no other way yet than dropping loopback device support, if
> >this
> >> approach is not acceptable. But, we found it nice to have, since
> >performance
> >> for the loopback case improved significantly.
> >>
> >> Any suggestion on how to enable loopback support (referencing
> >127.0.0.x
> >> addresses during connection management) in a more elegant way would
> >> be highly appreciated....
> >>
> >
> >RXE solved it by relying on loopback from netdevice, see
> >rxe_dma_device call
> >in rxe_register_device.
> >
> >Will it work for SIW?
> >
> >Thanks
> >
>
> Right...this is basically what siw can do as well for all
> other devices it attaches to. But it will not work for the
> real loopback device (addr 127.0.0.1 and friends), since that
> device does not come with a parent device it belongs
> to. Loopback seem to be kind of a 'top level' device.
>
> rxe_dma_device() returns netdev->dev.parent, which is
> NULL for loopback. So there is probably currently no support
> for lo devices within rxe.
>
> I'll take the loopback devices itself as the parent device.
> That seem to work. I'd suggest that for rxe as well.

Feel free to send a patch for RXE.

Thanks

>
> Thank you!
>

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