Re: [dpdk-dev] [PATCH 1/2] ethdev: add flow action type update as an offload

[Thomas Monjalon <thomas@monjalon.net>]

08/11/2019 12:40, Zhang, Qi Z:
> 
> > -----Original Message-----
> > From: dev <dev-bounces@dpdk.org> On Behalf Of Thomas Monjalon
> > Sent: Friday, November 8, 2019 7:04 PM
> > To: Andrew Rybchenko <arybchenko@solarflare.com>
> > Cc: Ori Kam <orika@mellanox.com>; dev@dpdk.org;
> > pbhagavatula@marvell.com; Yigit, Ferruh <ferruh.yigit@intel.com>;
> > jerinj@marvell.com; Mcnamara, John <john.mcnamara@intel.com>;
> > Kovacevic, Marko <marko.kovacevic@intel.com>; Adrien Mazarguil
> > <adrien.mazarguil@6wind.com>; david.marchand@redhat.com;
> > ktraynor@redhat.com; Olivier Matz <olivier.matz@6wind.com>
> > Subject: Re: [dpdk-dev] [PATCH 1/2] ethdev: add flow action type update as an
> > offload
> > 
> > 08/11/2019 11:42, Andrew Rybchenko:
> > > On 11/8/19 1:28 PM, Thomas Monjalon wrote:
> > > > 08/11/2019 09:35, Andrew Rybchenko:
> > > >> The problem:
> > > >> ~~~~~~~~~~~~
> > > >> PMD wants to know before port start if application wants to to use
> > > >> flow MARK/FLAG in the future. It is required because:
> > > >>
> > > >> 1. HW may be configured in a different way to reserve resources
> > > >>    for MARK/FLAG delivery
> > > >>
> > > >> 2. Datapath implementation choice may depend on it (e.g. vPMD
> > > >>    is faster, but does not support MARK)
> > > >
> > > > Thank you for the clear problem statement.
> > > > I agree with it. This is a real design issue.
> > > >
> > > >
> > > >> Discussed solutions:
> > > >> ~~~~~~~~~~~~~~~~~~~~
> > >
> > > May be it is not 100% clear since below are alternatives.
> > >
> > > >> A. Explicit Rx offload suggested by the patch.
> > > >>
> > > >> B. Implicit by validation of a flow rule with MARK/FLAG actions used.
> > > >>
> > > >> C. Use dynamic field/flag (i.e. application registers dynamic field
> > > >>    and/or flag and PMD uses lookup to solve the problem) plus part
> > > >>    of (B) to discover if a feature is supported.
> > > >
> > > > The dynamic field should be registered via a new API function named
> > > > '<feature>_init'.
> > > > It means the application must explicit request the feature.
> > > > I agree this is the way to go.
> > >
> > > If I understand your statement correctly, but (C) is not ideal since
> > > it looks global. If registered dynamic field of mbuf and is flag that
> > > the feature should be enabled, it is a flag to all ports/devices.
> > >
> > > >> All solutions require changes in applications which use these
> > > >> features. There is a deprecation notice in place which advertises
> > > >> DEV_RX_OFFLOAD_FLOW_MARK addition, but may be it is OK to
> > > >> substitute it with solution (B) or (C). Solution (C) requires
> > > >> changes since it should be combined with (B) in order to understand
> > > >> if the feature is supported.
> > > >
> > > > I don't understand.
> > > > Application request and PMD support are two different things.
> > > > PMD support must be via rte_flow validation on a case by case anyway.
> > >
> > > I mean that application wants to understand if the feature is
> > > supported. Then, it wants to enable it. In the case of (B), if I
> > > understand the solution correctly, there is no explicit way to enable,
> > > PMD just detects it because of discovery is done (that's what I mean
> > > by "implicit" and it is a drawback from my point of view, but still
> > > could be considered). (C) solves the problem of (B).
> > >
> > > >> Advantages and drawbacks of solutions:
> > > >> ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> > > >> 1. The main drawback of (A) is a "duplication" since we already
> > > >>    have a way to request flow MARK using rte_flow API.
> > > >>    I don't fully agree that it is a duplication, but I agree
> > > >>    that it sounds like duplication and complicates a bit flow
> > > >>    MARK usage by applications. (B) complicates it as well.
> > > >>
> > > >> 2. One more drawback of the solution (A) is the necessity of
> > > >>    similar solution for META and it eats one more offload bit.
> > > >>    Yes, that's true and I think it is not a problem.
> > > >>    It would make it easier for applications to find out if
> > > >>    either MARK or META is supported.
> > > >>
> > > >> 3. The main advantage of the solution (A) is simplicity.
> > > >>    It is simple for application to understand if it supported.
> > > >>    It is simple in PMD to understand that it is required.
> > > >>    It is simple to disable it - just reconfigure.
> > > >>    Also it is easier to document it - just mention that
> > > >>    the offload should be supported and enabled.
> > > >>
> > > >> 4. The main advantage of the solution (B) is no "duplication".
> > > >>    I agree that it is valid argument. Solving the problem
> > > >>    without extra entities is always nice, but unfortunately
> > > >>    it is too complex in this case.
> > > >>
> > > >> 5. The main drawback of the solution (B) is the complexity.
> > > >>    It is necessary to choose a flow rule which should be used
> > > >>    as a criteria. It could be hardware dependent.
> > > >>    Complex logic is require in PMD if it wants to address the
> > > >>    problem and control MARK delivery based on validated flow
> > > >>    rules. It adds dependency between start/stop processing and
> > > >>    flow rules validation code.
> > > >>    It is pretty complicated to document it.
> > > >>
> > > >> 6. Useless enabling of the offload in the case of solution (A)
> > > >>    if really used flow rules do not support MARK looks like
> > > >>    drawback as well, but easily mitigated by a combination
> > > >>    with solution (B) and only required if the application wants
> > > >>    to dive in the level of optimization and complexity and
> > > >>    makes sense if application knows required flow rules in
> > > >>    advance. So, it is not a problem in this case.
> > > >>
> > > >> 7. Solution (C) has drawbacks of the solution (B) for
> > > >>    applications to understand if these features are supported,
> > > >>    but no drawbacks in PMD, since explicit criteria is used to
> > > >>    enable/disable (dynamic field/flag lookup).
> > > >>
> > > >> 8. Solution (C) is nice since it avoids "duplication".
> > > >>
> > > >> 9. The main drawback of the solution (C) is asymmetry.
> > > >>    As it was discussed in the case of RX_TIMESTAMP
> > > >>    (if I remember it correctly):
> > > >>     - PMD advertises RX_TIMESTAMP offload capability
> > > >>     - application enables the offload
> > > >>     - PMD registers dynamic field for timestamp
> > > >>    Solution (C):
> > > >>      - PMD advertises nothing
> > > >>      - application uses solution (B) to understand if
> > > >>        these features are supported
> > > >>      - application registers dynamic field/flag
> > > >>      - PMD does lookup and solve the problem
> > > >>    The asymmetry could be partially mitigated if RX_TIMESTAMP
> > > >>    solution is changed to require an application to register
> > > >>    dynamic fields and PMD to do lookup if the offload is
> > > >>    enabled. So, the only difference will be in no offload
> > > >>    in the case of flow MARK/FLAG and usage of complex logic
> > > >>    to understand if it is supported or no.
> > > >>    May be it would be really good since it will allow to
> > > >>    have dynamic fields registered before mempool population.
> > > >>
> > > >> 10. Common drawback of solutions (B) and (C) is no granularity.
> > > >>     Solution (A) may be per queue while (B) and (C) cannot be
> > > >>     per queue. Moreover (C) looks global - for all devices.
> > > >>     It could be really painful.
> > > >>
> > > >> (C) is nice, but I still vote for simplicity and granularity of
> > > >> (A).
> > > >
> > > > I vote for clear separation of application needs and PMD support, by
> > > > using the method C (dynamic fields).
> > > > I agree timestamp must use the same path.
> > > > I agree it's complicate because we don't know in advance whether a
> > > > flow rule will be accepted, but that's the reality, config is complex.
> > >
> > > Do you think that global nature of the (C) is acceptable?
> > 
> > That's a good question.
> > Maybe the feature request should be per port.
> > In this case, we are back to solution A with a flag per port?
> > 
> > Note that A and C will not guarantee that the offload will be possible.
> > We need B (flow rule validation) anyway.
> 
> I may not understand how solution B can works well for all the cases.

I think you didn't read above carefully.
I am not saying B will solve all, but is needed in addition of A and C.

> A rte_flow rule can be issued after dev_start, which means the rx_burst function is already selected at that time, 
> so does that mean the driver need to switch from a non- mark offload aware path to a mark offload aware path without stop device?

I agree to have the application request the offload before starting (A).

> or it has to reject the flow? 

Yes if PMD is not ready (ignored app request or app did not request),
it must reject the flow rule.

> The question is if we have 2 data path, one support some offload , one not but more fast, which one should be selected during dev_start? Isn't Offload widely used to solve this problem?
> 
> I think the option A solve all the problems, option C might also works, but A is looks much straightforward for me.

Again, the answer is below:

> > It seems A, B, C are not alternatives but all required as pieces of a puzzle...