[dpdk-dev] [PATCH v2 1/2] ethdev: replace callback getting filter operations

[Thomas Monjalon]

20/03/2021 08:54, Andrew Rybchenko:
> On 3/19/21 9:12 PM, Thomas Monjalon wrote:
> > 15/03/2021 10:15, Thomas Monjalon:
> >> 15/03/2021 10:08, Andrew Rybchenko:
> >>> On 3/15/21 11:55 AM, Thomas Monjalon wrote:
> >>>> 15/03/2021 09:43, Andrew Rybchenko:
> >>>>> On 3/15/21 10:54 AM, Thomas Monjalon wrote:
> >>>>>> 15/03/2021 08:18, Andrew Rybchenko:
> >>>>>>> On 3/12/21 8:46 PM, Thomas Monjalon wrote:
> >>>>>>>> --- a/lib/librte_ethdev/rte_flow.c
> >>>>>>>> +++ b/lib/librte_ethdev/rte_flow.c
> >>>>>>>> @@ -255,18 +255,19 @@ rte_flow_ops_get(uint16_t port_id, struct rte_flow_error *error)
> >>>>>>>>   
> >>>>>>>>   	if (unlikely(!rte_eth_dev_is_valid_port(port_id)))
> >>>>>>>>   		code = ENODEV;
> >>>>>>>> -	else if (unlikely(!dev->dev_ops->filter_ctrl ||
> >>>>>>>> -			  dev->dev_ops->filter_ctrl(dev,
> >>>>>>>> -						    RTE_ETH_FILTER_GENERIC,
> >>>>>>>> -						    RTE_ETH_FILTER_GET,
> >>>>>>>> -						    &ops) ||
> >>>>>>>> -			  !ops))
> >>>>>>>> -		code = ENOSYS;
> >>>>>>>> +	else if (unlikely(dev->dev_ops->flow_ops_get == NULL))
> >>>>>>>> +		code = ENOTSUP;
> >>>
> >>> It is described as:
> >>>     -ENOTSUP: valid but unsupported rule specification (e.g.
> >>>     partial bit-masks are unsupported).
> >>> So, it looks different. May be it is really better to keep
> >>> ENOSYS.
> >>>
> >>>>>>>>   	else
> >>>>>>>> -		return ops;
> >>>>>>>> -	rte_flow_error_set(error, code, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
> >>>>>>>> -			   NULL, rte_strerror(code));
> >>>>>>>> -	return NULL;
> >>>>>>>> +		code = dev->dev_ops->flow_ops_get(dev, &ops);
> >>>>>>>> +	if (code == 0 && ops == NULL)
> >>>>>>>> +		code = EACCES;
> >>>>>>> It looks something new. I think it should be mentioned in flow_ops_get
> >>>>>>> type documentation (similar to eth_promiscuous_enable_t) and
> >>>>>>> rte_flow_validate() etc functions
> >>>>>>> return values description.
> >>>>>>
> >>>>>> It is an internal function used only in rte_flow.c.
> >>>>>> The real consequence is to set rte_errno in a lot of rte_flow API.
> >>>>>> Not sure there is a good way to document the code details.
> >>>>>> Other codes are not documented in rte_flow.h
> >>>>>
> >>>>> First of all it is a behaviour of the flow_ops_get callback and
> >>>>> driver developers should know that it is a legal to return 0 and
> >>>>> ops==NULL and know what it means.
> >>>>
> >>>> The combination code 0 and ops NULL is not new.
> >>>> Previously, it was returning ENOSYS.
> >>>> I've just given a more meaningful error code: EACCES,
> >>>> while replacing ENOSYS with ENOTSUP for the other case.
> >>>
> >>> Yes, exactly. What I'm trying to say that it would be
> >>> helpful to make it a bit more transparent to PMD developers.
> >>> Yes, it was not documented before, I agree. I think it is
> >>> a good time to improve documentation.
> >>>
> >>>>> Second, it is visible as rte_flow_validate() (and other functions
> >>>>> which use rte_flow_ops_get()) return value value which has
> >>>>> special meaning. So, should be documented.
> >>>>
> >>>> Yes, I should update the API doc where ENOSYS was mentioned.
> >>>> Or probably better: I should keep the error code ENOSYS
> >>>> and do not break API.
> >>>> Preference?
> >>>
> >>> Good question. I think we should not distinguish NULL callback
> >>> and NULL ops returned by not-NULL callback. So, I think
> >>> keeping ENOSYS is the best option here.
> >>
> >> OK, thank you for the review.
> >> So the conclusion is: keep ENOSYS and document NULL ops case.
> > 
> > After more thoughts, I don't think we need to insist on the NULL ops case.
> > The function rte_flow_ops_get returns the ops,
> > and it is documented that returning NULL is an error.
> > So the function is just setting ENOSYS error code to have
> > an error code associated with returning NULL.
> > For the PMD, returning an ops NULL has no interest.
> > 
> 
> May be I misunderstand above, but
> 
> One driver may support different NIC types. It could support flow API
> for one NIC type and do not support for another. Sometimes it could
> be easier to sort it out in flow ops get callback rather than
> provide different set of ethdev ops callback. If so, returning NULL
> ops makes sense and easier way to say that flow API is not supported
> (vs, for example, providing dummy callbacks which return ENOSYS).

Yes you're right, returning NULL ops may have an interest.
I will document it in eth_flow_ops_get_t.