Re: [dpdk-dev] [RFC] ethdev: support input set change by RSS action

[Adrien Mazarguil <adrien.mazarguil@6wind.com>]

On Thu, Jul 04, 2019 at 01:55:14PM +0000, Zhang, Qi Z wrote:
> 
> 
> > -----Original Message-----
> > From: Adrien Mazarguil [mailto:adrien.mazarguil@6wind.com]
> > Sent: Thursday, July 4, 2019 5:07 PM
> > To: Su, Simei <simei.su@intel.com>
> > Cc: Zhang, Qi Z <qi.z.zhang@intel.com>; Wu, Jingjing <jingjing.wu@intel.com>;
> > Xing, Beilei <beilei.xing@intel.com>; Yang, Qiming <qiming.yang@intel.com>;
> > dev@dpdk.org
> > Subject: Re: [dpdk-dev] [RFC] ethdev: support input set change by RSS action
> > 
> > On Thu, Jul 04, 2019 at 12:47:09PM +0800, simei wrote:
> > > From: Simei Su <simei.su@intel.com>
> > >
> > > This RFC introduces inputset structure to rte_flow_action_rss to
> > > support input set specific configuration by rte_flow RSS action.
> > >
> > > We can give an testpmd command line example to make it more clear.
> > >
> > > For example, below flow selects the l4 port as inputset for any
> > > eth/ipv4/tcp packet: #flow create 0 ingress pattern eth / ipv4 / tcp /
> > > end actions rss inputset tcp src mask 0xffff dst mask 0xffff /end
> > >
> > > Signed-off-by: Simei Su <simei.su@intel.com>
> > > ---
> > >  lib/librte_ethdev/rte_flow.h | 3 +++
> > >  1 file changed, 3 insertions(+)
> > >
> > > diff --git a/lib/librte_ethdev/rte_flow.h
> > > b/lib/librte_ethdev/rte_flow.h index f3a8fb1..2a455b6 100644
> > > --- a/lib/librte_ethdev/rte_flow.h
> > > +++ b/lib/librte_ethdev/rte_flow.h
> > > @@ -1796,6 +1796,9 @@ struct rte_flow_action_rss {
> > >  	uint32_t queue_num; /**< Number of entries in @p queue. */
> > >  	const uint8_t *key; /**< Hash key. */
> > >  	const uint16_t *queue; /**< Queue indices to use. */
> > > +	struct rte_flow_item *inputset; /** Provide more specific inputset
> > configuration.
> > > +					 * ignore spec, only mask.
> > > +					 */
> > >  };
> > >
> > >  /**
> > 
> > To make sure I understand, is this kind of a more flexible version of
> > rte_flow_action_rss.types?
> 
> Yes
> > 
> > For instance while specifying .types = ETH_RSS_IPV4 normally covers both
> > source and destination addresses, does this approach enable users to perform
> > RSS on source IP only? 
> 
> Yes, .it is the case to select any subset of 5 tuples or even tunnel header's id for hash
> 
> > In which case, what value does the Toeplitz algorithm
> > assume for the destination, 0x0? (note: must be documented)
> 
> My understanding is src/dst pair is only required for a symmetric case
> But for Toeplitz, it is just a hash function, it process a serial of data with specific algorithm, have no idea about which part is src and dst , 
> So for ip src only with Toeplitz, dst is not required to be a placeholder..

Right, I had symmetric Toeplitz in mind and wondered what would happen when
users would not select the required fields. I guess the PMD would have to
reject unsupported combinations.

> anything I missed, would you share more insight?

No, that answers the question, thanks.

Now what about my suggestion below? In short: extending ETH_RSS_* assuming
there's enough bits left in there, instead of adding a whole new framework
and breaking ABI in the process.

> > My opinion is that, unless you know of a hardware which can perform RSS on
> > random bytes of a packet, this approach is a bit overkill at this point.
> > 
> > How about simply adding the needed ETH_RSS_* definitions (e.g.
> > ETH_RSS_IPV4_(SRC|DST))? How many are needed?
> > 
> > There are currently 20 used bits and struct rte_flow_action_rss.types is 64-bit
> > wide. I'm sure we can manage something without harming the ABI. Even
> > better, you wouldn't need a deprecation notice.
> > 
> > If you use the suggested approach, please update testpmd and its
> > documentation as part of the same patch, thanks.

-- 
Adrien Mazarguil
6WIND