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

[Thomas Monjalon <thomas@monjalon.net>]

08/11/2019 14:27, Andrew Rybchenko:
> On 11/8/19 4:17 PM, Thomas Monjalon wrote:
> > 08/11/2019 13:00, Andrew Rybchenko:
> >> On 11/8/19 2:03 PM, Thomas Monjalon wrote:
> >>> 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?
> >> Offloads are natively per-queue as well, so (A) keeps the choice
> >> between per-port vs per-queue to PMDs as usual.
> >>
> >>> Note that A and C will not guarantee that the offload will be possible.
> >> Yes, definitely.
> >>
> >>> We need B (flow rule validation) anyway.
> >> Strictly speaking (B) (checking flow rules before device
> >> startup) is required if an application can predict flow
> >> rules and wants to ensure that MARK offload will be usable.
> >> Otherwise, it may be skipped.
> > No no, I mean flow rule validation MUST be used anyway
> > during the runtime before applying a rule.
> > I agree it is hard to predict. I speak only about real rules.
> 
> OK, I see. Of course, flow rule validation is required at runtime.
> I was rather concentrated on the stated problem solutions.
> 
> >>> It seems A, B, C are not alternatives but all required
> >>> as pieces of a puzzle...
> >> Unfortunately true in the most complex case.
> >> Right now it will be A with B if required as explained above.
> >> C will come a bit later when the field migrates to dynamic.
> >>
> >> May be it is even better if application registers dynamic
> >> fields before an attempt to enable offload to be sure that
> >> it will not fail because of impossibility to register
> >> dynamic field (lack of space). I'm not sure, but it is not
> >> not that important.
> > Yes of course, lack of mbuf space is another reason for
> > disabling the feature.
> >
> >> If we finally go way A, should we add offloads for META back?
> >> I guess separate Rx and Tx are required.
> > I would prefer to add it as dynamic flags.
> > Why rushing on a very temporary solution while it is not a new issue?
> 
> Basically it means that we go just (B)+(C) in the case of META.
> I have no strong opinion but thought that it could be better to
> align the solution. Of course, we can wait with it. As I understand
> META is an experimental feature.

Yes it is experimental and I think it is too late to align now.

Anyway, we will probably to discuss again these offloads TAG/MARK/META,
as requested by several people.