Re: [RFC] ethdev: add send to kernel action

[Thomas Monjalon <thomas@monjalon.net>]

13/09/2022 14:09, Michael Savisko:
> From: Andrew Rybchenko <andrew.rybchenko@oktetlabs.ru>
> > On 9/12/22 16:39, Michael Savisko wrote:
> > > From: Thomas Monjalon <thomas@monjalon.net>
> > >> 16/08/2022 11:50, Ferruh Yigit:
> > >>> On 8/11/2022 12:35 PM, Michael Savisko wrote:
> > >>>>
> > >>>> In some cases application may receive a packet that should have
> > >>>> been received by the kernel. In this case application uses KNI or
> > >>>> other means to transfer the packet to the kernel.
> > >>>> This commit introduces rte flow action that the application may use
> > >>>> to route the packet to the kernel while still in the HW.
> > >>>>
> > >>>> Signed-off-by: Michael Savisko <michaelsav@nvidia.com>
> > >>>
> > >>> I assume this only works for bifurcated drivers, right?
> > >>
> > >> This question has not been replied after a month.
> > >> Please let's be more reactive.
> >  >
> >  > Depends on HW. If it can forward packets to different places then it can also
> > be supported. But in most cases yes - for bifurcated drivers.
> > 
> > The action sounds like "do some magic". As far as I know we have no concept of
> > kernel and cooperation with the kernel in DPDK yet.
> 
> There's nothing "magical". Kernel is not a part of DPDK, but DPDK can use KNI to transfer messages between application and kernel.
> With bifurcated driver we can have a rule to route the packet matching a pattern (example: IPv4 packets) to the DPDK application and the rest of the traffic will be received by the kernel. 
> But if we want to receive most of the traffic in DPDK except specific pattern (example: ICMP packets) that should be processed by the kernel, then it's easier to re-route these packets with a single rule.
> The new action I'm suggesting allows application to route packets directly to the kernel without software involvement, it is a HW offload.
> We see it used when working with bifurcated driver, because the kernel driver and the DPDK driver are sharing the same HW.
> 
> > Is it a transfer or non-transfer action?
> > I guess non-transfer, since otherwise the next question is which kernel...
> 
> This is an ingress action only.

Should we add this note in the doxygen comment?
This is the wording in the v2 sent today:

+       /*
+        * Send packets to the kernel, without going to userspace at all.
+        * The packets will be received by the kernel driver sharing
+        * the same device as the DPDK port.
+        *
+        * No associated configuration structure.
+        */
+       RTE_FLOW_ACTION_TYPE_SEND_TO_KERNEL,

> > In the case of non-transfer DPDK has a concept of Rx queues which are used to
> > deliver traffic to and we have QUEUE and RSS flow actions to do it.
> 
> The idea of this offload action is to route traffic away from the DPDK application.
> 
> > The patch adds some magic direction "kernel". Don't we want to control
> > destination queue? RSS?
> > May be we need dedicated control steps to setup kernel Rx queues and than use
> > QUEUE/RSS to direct traffic there?
> 
> We have no control of how the kernel is configured.