Re: [dpdk-dev] [PATCH] doc: remove deprecated ethdev features

[Thomas Monjalon <thomas@monjalon.net>]

16/10/2019 12:20, Jerin Jacob:
> On Wed, 16 Oct, 2019, 3:46 PM Ferruh Yigit, <ferruh.yigit@intel.com> wrote:
> 
> > On 10/16/2019 11:08 AM, Jerin Jacob wrote:
> > >
> > >
> > > On Wed, 16 Oct, 2019, 3:32 PM Ferruh Yigit, <ferruh.yigit@intel.com
> > > <mailto:ferruh.yigit@intel.com>> wrote:
> > >
> > >     On 10/15/2019 5:19 PM, Jerin Jacob wrote:
> > >     > On Tue, Oct 15, 2019 at 9:26 PM Ferruh Yigit <
> > ferruh.yigit@intel.com
> > >     <mailto:ferruh.yigit@intel.com>> wrote:
> > >     >>
> > >     >> On 10/15/2019 3:16 PM, Jerin Jacob wrote:
> > >     >>>>>>> @@ -36,13 +36,6 @@ VMDq                 =
> > >     >>>>>>>   SR-IOV               =
> > >     >>>>>>>   DCB                  =
> > >     >>>>>>>   VLAN filter          =
> > >     >>>>>>> -Ethertype filter     =
> > >     >>>>>>> -N-tuple filter       =
> > >     >>>>>>> -SYN filter           =
> > >     >>>>>>> -Tunnel filter        =
> > >     >>>>>>> -Flexible filter      =
> > >     >>>>>>> -Hash filter          =
> > >     >>>>>>> -Flow director        =
> > >     >>>>>>>   Flow control         =
> > >     >>>>>>>   Flow API             =
> > >     >>>>>>>   Rate limitation      =
> > >     >>>>>> I suggest adding these features back!
> > >     >>>>>>
> > >     >>>>>> "Flow director" and other filters are features that
> > device/driver
> > >     supports.
> > >     >>>>>>
> > >     >>>>>> And "Flow API" and "filter_ctrl" are methods used to
> > implement these
> > >     features.
> > >     >>>>>> Indeed they are only different APIs to get input from
> > application/user.
> > >     >>>>>>
> > >     >>>>>> It doesn't really mean much to say "Flow API" is supported?
> > So what
> > >     is really
> > >     >>>>>> supported? It matters more what feature is supported.
> > >     >>>>>>
> > >     >>>>>> Since we are saying old method is deprecated, we can update
> > the
> > >     feature list of
> > >     >>>>>> drivers which implements filtering features using old method
> > as not
> > >     supported.
> > >     >>>>>> And that is the case with this patch since old APIs are
> > marked as
> > >     deprecated,
> > >     >>>>>> users can't use them to enable a filtering feature.
> > >     >>>>>>
> > >     >>>>>> Indeed I am for removing the "Flow API" from feature list,
> > first it
> > >     is not a
> > >     >>>>>> feature, second if it is only method to enable a filtering,
> > and if
> > >     filtering is
> > >     >>>>>> enabled in a driver, what is the point of redundant "Flow
> > API" listing?
> > >     >>>>>>
> > >     >>>>>> I can make a quick patch if there is no objection, thanks.
> > >     >>>>>
> > >     >>>>> As I understand it was a decision to avoid details about flow
> > API support
> > >     >>>>> in features matrix. Mainly because matrix would be really huge
> > in
> > >     >>>>> attempt to represent it. The question is why filters/patterns
> > mentioned
> > >     >>>>> above are better than others and should be mentioned.
> > >     >>>>> I'm not against adding some details, just want to understand
> > criteria.
> > >     >>>>> Flexible and hash are definitely not well defined.
> > >     >>>>> What is flow director and which features should be supported
> > to say yes?
> > >     >>>>>
> > >     >>>
> > >     >>>>
> > >     >>>> The criteria I have is what users will be interested about a
> > device/driver.
> > >     >>>>
> > >     >>>> Will it be really huge to list filtering capabilities of the
> > devices? I
> > >     believe
> > >     >>>> we can group them into a few groups like above.
> > >     >>>> Or at least keep existing one and improve it by time and +1 to
> > clarify
> > >     them more
> > >     >>>> but that is something else.
> > >     >>>>
> > >     >>>> A device can have capabilities but it is not easy to find if
> > that
> > >     capability has
> > >     >>>> been enabled via DPDK, this kind of feature matrix works for
> > it, and all
> > >     >>>> features together makes it much easier than digging datasheets
> > and code.
> > >     >>>>
> > >     >>>> Saying that "Flow API" is enabled for a driver doesn't really
> > gives any
> > >     >>>> information to the user if they are interested what kind of
> > filtering
> > >     features
> > >     >>>> are supported by that device/driver.
> > >     >>>
> > >     >>> I agree. I think, we need to enumerate rte flow patterns and
> > actions
> > >     >>> supported by the PMD.
> > >     >>> Since there was no single place such documentation, we added the
> > same
> > >     >>> in PMD documentation
> > >     >>> See 39.8. RTE Flow Support at
> > >     https://doc.dpdk.org/guides/nics/octeontx2.html
> > >     >>>
> > >     >>> And IMO, We should not add deprecated features in the features
> > matrix as
> > >     >>> new PMDs are not planning to implement the deprecated APIs. That
> > >     >>> makes, matrix looks
> > >     >>> new PMDs do not implement a lot of features, but in reality,
> > those are
> > >     >>> deprecated and never planning to implement if even though HW
> > supports it.
> > >     >>>
> > >     >>
> > >     >> +1 to not add deprecated features to the matrix, but those
> > removed ones
> > >     [1] are
> > >     >> not deprecated. Implementing them via "filter_ctrl" is
> > deprecated. Below
> > >     >> features still can be implemented via "Flow API", that is why I
> > am for adding
> > >     >> them back to default.ini.
> > >     >
> > >     > Got it. Instead of [1], Can we document it as in the form of
> > rte_flow
> > >     > semantics(patterns and actions) so
> > >     > that for the end-user it is very clear. Reason being:
> > >     > # Expressing "Tunnel filter" or "N-tupe filter" or "Flexible
> > filter"
> > >     > or "Flow director" etc is very vague in rte_flow semantics
> > >     > and function is not just limited with above-fixed functions
> > >     > #  The new PMDs also can express the rte_flow aka "Flow API"
> > support
> > >     > in the rte_flow semantics.
> > >
> > >     rte_flow is implementation detail, as well as 'filter_ctrl', I
> > believe listing
> > >     rte_flow semantic will be too much detail for the feature table.
> > >
> > >     And end user may be interested in features, as if that drive/device
> > support
> > >     "Flow Director" or not, instead of rte_flow semantic.
> > >
> > >     But I can see feature being vague is also problem, perhaps we can
> > have rte_flow
> > >     level details in features.rst file, will it work?
> > >
> > >
> > >
> > > +1 for adding rte_flow level level details in features.rst
> >
> > OK, let me check this
> >
> 
> Ok

Just to confirm my thought:
The name of the removed "features" came from the filter API,
which is very similar to Intel datasheets.
In rte_flow API, these categories may not make sense.
I am OK to not refer to rte_flow but to show the real implemented features,
if you find a way to express them concisely.


> > > IMO, Supported packet types(ptype) also good addition in features list.
> >
> > "Packet type parsing" feature is already there,
> >
> > http://lxr.dpdk.org/dpdk/v19.08/source/doc/guides/nics/features/default.ini#L53
> >
> > If you mean the list of supported types, it is possible to get list on
> > runtime
> > via an API, it will be hard to maintain that list in documentation.
> >
> 
> Yes. I meant the list of supported types.
> Ok. I will check the feasibility.

We need to be careful about the size of this matrix.
We can split in several matrix if needed.