Re: [dpdk-dev] [PATCH v2 3/3] ethdev: improve flow mark Rx offload deprecation notice

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 12/16/19 10:38 AM, Jerin Jacob wrote:
> On Mon, Dec 9, 2019 at 2:47 PM Andrew Rybchenko
> <arybchenko@solarflare.com> wrote:
>>
>> On 12/5/19 11:12 AM, Jerin Jacob wrote:
>>> On Mon, Dec 2, 2019 at 5:27 PM Andrew Rybchenko
>>> <arybchenko@solarflare.com> wrote:
>>>> 
>>>>>>
>>>>> Ack.
>>>>
>>>> Yes, I agree as well, but in general we already have an
>>>> exception MBUF_FAST_FREE which is just a nice wrap for
>>>> enabled by default support for mbufs from different
>>>> mempools and support for mbuf reference counters.
>>>> I don't suggest to change it. Just want to highlight
>>>> that we already have exceptions (nicely wrapped).
>>>> That's why I would not touch packet type parsing
>>>> "offload". Packet type parsing is not just on/off and
>>>> adding on/off in addition to existing API looks overkill.
>>>> Yes, it is one more exception, but nicely wrapped as well.
>>>
>>> I am all for making offloads disabled by default.
>>>
>>>>
>>>>>>> And what would be DEFAULT behavior for the mbuf MARK updation feature?
>>>>>>> (That's where this thread started).
>>>>>>
>>>>>> As all other features, mark is disabled by default.
>>>>>> Using a rte_flow rule, it can be enabled.
>>>>>> No need to pre-enable it.
>>>>>
>>>>> Ok.
>>>>
>>>> But it returns us to the point where we started [1]:
>>>>
>>>> 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).
>>>>
>>>> opt-in/opt-out solution has drawbacks mentioned above.
>>>> Also I'm not sure if opt-in/opt-out is per-queue or per-port.
>>>> (Offloads may be naturally per-queue and it is a big advantage).
>>>>
>>>> IMHO feature which should be opt-out is almost equivalent to
>>>> offload enabled by default. It has the same negative properties
>>>> as enabled by default offloads.
>>>>
>>>> Am I missing something again?
>>>>
>>>> From my point of view I see no problem in necessity to say
>>>> in advance (before device start) that application would like
>>>> to use some features at run time.
>>>
>>> I agree with your problem definition and solution as offload.
>>>
>>> I think, our constraint is, we can not change functional ABI behavior
>>> for the next year. i.e The existing application should work for the
>>> next year without
>>> changing the code.
>>>
>>> I think, it all boiling down to adhere to that constraint or not for
>>> this specific case.
>>
>> May be the escape is to avoid consistency checks in generic
>> code (not sure that such checks are doable/required in this
>> case, but anyway) and make the behaviour change vendor/driver-
>> specific. I understand that it is far from ideal solution.
>>
>> May be offload should be combined with opt-out as a way to
>> disable. I.e. offload is positive (not negative), but enabled
>> by default (i.e. automatically added to offloads as we do
>> for RSS_HASH) with an experimental opt-out to disable it.
>>
>> As the result:
>> 1. There is no changes in behaviour from application point of
>>    view.
>> 2. Application which care about performance and ready to use
>>    experimental opt-out to optimize performance can do it.
>>    (i.e. use opt-out to avoid the offload enabled by default).
>> 3. Later when window to normalize behaviour opens, opt-out
>>    becomes NOP (i.e. it still could be preserved for some
>>    time to simplify transition).
>> 4. The offload is enabled by default during transition
>>    period only since it represents a feature which had
>>    no offload flag before and was always enabled before.
>> 5. As an offload the feature may be controlled per-device
>>    and per-queue natively.
> 
> Looks good to me.
> It makes sense to have a generic opt API to have for year ABI,
> which works on
> 
> - per queue/per port
> - Enable by default to keep backward compatible.
> - Have a generic signature to allow probe() all the enabled opt-in features
> and then disable if required by the application.

I'd like to clarify to be sure that we're on the same page:
1. Add DEV_RX_OFFLOAD_FLOW_MARK offload:
   - enabled by default till 20.11 to preserve behaviour
   - applications may migrate and explicitly enable
   - disabled by default since 20.11 to switch to generic
     policy which require offloads to be disabled by default
2. Add experimental opt-out which allow to disable the
   offload to optimize performance for applications which
   would like to care about it early
   - opt-out remains but becomes NOP in 20.11

> - I think, rte_eth_dev_set_ptypes()  needs to change to generic API as
> it comes under opt-in/out scheme.

I'm not sure that I understand how it should look like for
ptypes.

>>
>> It still does not sort out "necessity to enable twice"
>> concern which for specified above "the problem", IMO,
>> contradicts to "disabled by default offloads" (I read
>> it as "the best performance" by default).
>>
>>> Once that is decided, we can wrap it in offload flags vs opt scheme
>>> (by default enabled scheme).
>>
>> Yes. May be I don't understand all the details of the opt
>> scheme right now, but I don't like what I can imagine as
>> described above.
>>
>>>>
>>>> Yes, all features which may be controlled at run-time are
>>>> headache for optimizations (VLAN offloads).
>>>>
>>>> [1]
>>>> http://inbox.dpdk.org/dev/f170105b-9c60-1b04-cb18-52e0951ddcdb@solarflare.com/
>>