[dpdk-dev] [PATCH v9 12/12] abi: fix v2.1 abi broken issue

[Liang, Cunming]

Hi Stephen,

On 6/1/2015 9:27 PM, Stephen Hemminger wrote:
> On Mon, 1 Jun 2015 16:48:01 +0800
> "Liang, Cunming" <cunming.liang at intel.com> wrote:
>
>> Hi Stephen,
>>
>> On 5/29/2015 11:27 PM, Stephen Hemminger wrote:
>>> On Fri, 29 May 2015 16:45:25 +0800
>>> Cunming Liang <cunming.liang at intel.com> wrote:
>>>   
>>>> +#ifdef RTE_EAL_RX_INTR
>>>> +extern int
>>>>    rte_eth_dev_rx_intr_ctl(uint8_t port_id, int epfd, int op, void *data);
>>>> +#else
>>>> +static inline int
>>>> +rte_eth_dev_rx_intr_ctl(uint8_t port_id, int epfd, int op, void *data)
>>>> +{
>>>> +	RTE_SET_USED(port_id);
>>>> +	RTE_SET_USED(epfd);
>>>> +	RTE_SET_USED(op);
>>>> +	RTE_SET_USED(data);
>>>> +	return -1;
>>>> +}
>>>> +#endif
>>> Doing ABI compatibility is good but hard.
>>>
>>> I think it would be better not to provide the functions for rx_intr_ctl unless
>>> the feature was configured on. That way anyone using them with incorrect config
>>> would detect failure at build time, rather than run time.
>> I tend to not agree. For rx_intr_ctl/rx_intr_ctl_q, no matter w/ or w/o
>> RTE_EAL_RX_INTR, it's necessary to check the return value.
>> The failure return shall cause application give up using epoll waiting
>> on the specified epfd for the port, and then degraded to pure polling mode.
>> So I think these failure should be handled by the caller.
> It is always best to fail as early in the development process as possible.
> What possible benefit could there be from allowing application to be linked
> and run with incorrect configuration.
In fact you'll always detect failure at build time if your application 
insist to call rx_intr_ctl with port_conf.intr_conf.rxq=1.
As port_conf.intr_conf.rxq is not defined when RTE_EAL_RX_INTR is not 
defined.
If you ignore port_conf.intr_conf.rxq, no matter RTE_EAL_RX_INTR is 
defined or not, you always will get failure return when calling rx_intr_ctl.
So I think the behavior doesn't break 'fail as early in the development 
process as possible'.
And I'd like to expose all new APIs in this version, if we don't provide 
this API, what about rte_intr_rx_ctl or others? They're probably used by 
other user application as well.