Re: [dpdk-dev] [PATCH v2 1/3] ethdev: add new RSS types for IPv6 prefix

[Thomas Monjalon <thomas@monjalon.net>]

08/07/2020 13:10, Zhang, Qi Z:
> From: Thomas Monjalon <thomas@monjalon.net>
> > 08/07/2020 11:45, Zhang, Qi Z:
> > > On 2020/7/7 19:06, Thomas Monjalon wrote:
> > > > 16/06/2020 10:16, Junfeng Guo:
> > > >> This patch defines new RSS offload types for IPv6 prefix with 32,
> > > >> 48,
> > > >> 64 bits of both SRC and DST IPv6 address.
> > > >>
> > > >> Signed-off-by: Junfeng Guo <junfeng.guo@intel.com>
> > > >> ---
> > > >>   lib/librte_ethdev/rte_ethdev.h | 51
> > ++++++++++++++++++++++++++++++++++
> > > >>   1 file changed, 51 insertions(+)
> > > >>
> > > >> diff --git a/lib/librte_ethdev/rte_ethdev.h
> > > >> b/lib/librte_ethdev/rte_ethdev.h index 631b146bd..5a7ba36d8 100644
> > > >> --- a/lib/librte_ethdev/rte_ethdev.h
> > > >> +++ b/lib/librte_ethdev/rte_ethdev.h
> > > >> @@ -538,6 +538,9 @@ struct rte_eth_rss_conf {
> > > >>   #define ETH_RSS_L4_DST_ONLY        (1ULL << 60)
> > > >>   #define ETH_RSS_L2_SRC_ONLY        (1ULL << 59)
> > > >>   #define ETH_RSS_L2_DST_ONLY        (1ULL << 58)
> > > >> +#define ETH_RSS_L3_PRE32           (1ULL << 57)
> > > >> +#define ETH_RSS_L3_PRE48           (1ULL << 56)
> > > >> +#define ETH_RSS_L3_PRE64           (1ULL << 55)
> > > >
> > > > PRE32, 48 and 64 are not obvious.
> > > > Why is it needed?
> > >
> > > there is typical usage for NAT64, which use 32 bit prefix for IPv6
> > > addresses, in this case flows over IPv4 and IPv6 will result in the
> > > same hash value, as well as 48, 64, which also have some corresponding
> > > use cases,
> > > > At least, please add comments for the values of this API.
> > >
> > > sure, we will add more comments.
> > > > Do we want to continue with the RTE_ prefix missing?
> > > > Can't we add the prefix for the new values?
> > 
> > I think you misunderstood this question. I am asking to change the name
> > ETH_RSS_L3_PRE32 to RTE_ETH_RSS_L3_PRE32
> 
> OK, we are going change all the ETH_RSS_xxx to RTE_ETH_RSS_xxx, or just the new values?
> the first option looks make sense to me.

You cannot break compatibility with the existing values,
but you can provide an alias to preserve compatibility.

> > > 32, 48, 64 are typical usage, and consider suffix pair we may add
> > > later, it will cost 6 bits so far we still have 27 bit left,  so it
> > > looks like will not be a problem in following couple releases.
> > 
> > Having some space left is not a reason to waste it :) If I understand well,
> > there is no standard for this API.
> > You are assigning some bits to some usage.
> > I don't find it generic and flexible enough.
> 
> Actually IPv6 address prefix is in spec, please check below RFC.
> https://tools.ietf.org/html/rfc6052#page-5

Quoting the RFC:
"
   the prefix shall be either the "Well-Known Prefix"
   or a "Network-Specific Prefix" unique to the organization
   deploying the address translators.
   The prefixes can only have one of the following lengths:
   32, 40, 48, 56, 64, or 96.
   (The Well-Known Prefix is 96 bits long, and can only be used
   in the last form of the table.)
"

So 40 and 56 are missing.

> So probably we are not wasting bits here, since this is a typical usage
> that DPDK can provide.
> Of cause more description is needed in the code here.
> 
> > If you want to limit the size of the match, we should have a generic syntax to
> > choose how many bits of the IPv6 address are taken into account for RSS. Or
> > maybe an IPv6 mask.
> 
> Yes, I believe at some moment, a more generic solution is mandatory,
> And I think that will not work if we stick on the 64 bits, new API need to be introduced and old one should be abandoned 
> 
> > 
> > > but anyway use 64 bits to represent RSS inputset can't meet the coming
> > > complex RSS usage, we may need to figure out some new APIs and
> > abandon
> > > the old one.
> > > A stacked protocol layer with bit field selector in each layer is
> > > under consideration, hope we can contribute some RFC at some moment.
> > > also feel free let us know your thought.
> > 
> > My thought is to discuss how to fit this need in future and avoid adding few
> > bits of temporary workaround.
> > API definition is serious and we must avoid temporary half solutions.