[dpdk-dev] [PATCH v2 0/2] ethdev: add support for raw flow type for flow director

Rybalchenko, Kirill kirill.rybalchenko at intel.com
Thu Oct 5 11:09:32 CEST 2017


Hi Thomas,

As Ferruh rightly said, this feature does not affect any
other drivers except i40e.
It was implemented as a temporary measure to accelerate
adoption of new protocols in DPDK.
It does not eliminate importance of rte_flow.
As you can see, we've already added GTP as a part of rte_flow.
(see Beilei Xing patch http://dpdk.org/ml/archives/dev/2017-October/077483.html)

Regards,
Kirill.

> -----Original Message-----
> From: Yigit, Ferruh
> Sent: Wednesday 4 October 2017 20:47
> To: Thomas Monjalon <thomas at monjalon.net>; Rybalchenko, Kirill
> <kirill.rybalchenko at intel.com>
> Cc: dev at dpdk.org; Chilikin, Andrey <andrey.chilikin at intel.com>; Xing, Beilei
> <beilei.xing at intel.com>; Wu, Jingjing <jingjing.wu at intel.com>
> Subject: Re: [dpdk-dev] [PATCH v2 0/2] ethdev: add support for raw flow
> type for flow director
> 
> On 10/4/2017 6:56 PM, Thomas Monjalon wrote:
> > 04/10/2017 19:44, Ferruh Yigit:
> >> On 10/4/2017 5:57 PM, Thomas Monjalon wrote:
> >>> 03/10/2017 21:02, Ferruh Yigit:
> >>>> On 9/20/2017 9:42 AM, Kirill Rybalchenko wrote:
> >>>>> For complex packets use raw flow type with pre-constructed packet
> >>>>> buffer instead of creating a packet internally in PMD.
> >>>
> >>> Sorry for not catching this series before.
> >>>
> >>> As it has been said several times on this mailing list, the flow
> >>> director API is deprecated.
> >>> I think everybody here knows that it is going to be replaced by
> >>> rte_flow.
> >>>
> >>> That's why it does not make sense to extend flow director.
> >>> We are not going to update PMDs to support a new type of legacy flow
> >>> director.
> >>> Please focus your efforts on rte_flow.
> >>
> >> As far as I can see this is not to to extend flow director. But
> >> driver uses this struct and adding a new feature into driver requires
> >> update in this struct.
> >>
> >> I guess idea was for new filter functionalities PMD should use
> >> rte_flow, that is the new shiny method we have, I see the point here.
> >> But I don't see the point of making these structs under use completely
> immutable.
> >
> > I don't know what is RTE_ETH_FLOW_RAW.
> > Let's start by explaining it, how it is used by users, and why this
> > struct is needed.
> 
> Let me answer as much as I get from patches, if something is missing or
> wrong Kirill needs to correct it.
> 
> Driver (i40e) works with static pre-defined flow and pctypes. But new
> feature DDP lets loading custom profiles and work with custom and perhaps
> not yet defined flow and pctypes. So there are a few other patches to make
> driver more dynamic.
> 
> For this case I40E_VALID_FLOW() fails with custom flow types, and
> RTE_ETH_FLOW_RAW used as kind of (void *) to be able to work with new
> dynamic types.
> 
> > Thanks
> >



More information about the dev mailing list