Re: [PATCH] cryptodev: mark experimental state

[Panu Matilainen <pmatilai@redhat.com>]

On 11/26/2015 03:51 PM, Doherty, Declan wrote:
>> -----Original Message-----
>> From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
>> Sent: Thursday, November 26, 2015 10:09 AM
>> To: Panu Matilainen; Doherty, Declan
>> Cc: dev@dpdk.org
>> Subject: Re: [dpdk-dev] [PATCH] cryptodev: mark experimental state
>>
>> 2015-11-26 10:00, Panu Matilainen:
>>> On 11/26/2015 09:39 AM, Panu Matilainen wrote:
>>>> On 11/25/2015 07:38 PM, Thomas Monjalon wrote:
>>>>> --- a/config/common_linuxapp
>>>>> +++ b/config/common_linuxapp
>>>>> @@ -319,6 +319,7 @@ CONFIG_RTE_PMD_PACKET_PREFETCH=y
>>>>>
>>>>>    #
>>>>>    # Compile generic crypto device library
>>>>> +# EXPERIMENTAL: API may change without prior notice
>>>>>    #
>>>>>    CONFIG_RTE_LIBRTE_CRYPTODEV=y
>>>>>    CONFIG_RTE_LIBRTE_CRYPTODEV_DEBUG=n
>>>> [...]
>>>>
>>>> I think an experimental library which declares itself exempt from the
>>>> ABI policy should not be compiled by default. That way anybody wanting
>>>> to try it out will be forced to notice the experimental status.
>>>>
>>>> More generally / longer term, perhaps there should be a
>>>> CONFIG_RTE_EXPERIMENTAL which wraps all experimental features and
>>>> defaults to off.
>>>
>>> On a related note, librte_mbuf_offload cannot be built if
>>> CONFIG_RTE_LIBRTE_CRYPTODEV is disabled. Which seems to suggest its (at
>>> least currently) so tightly couple to cryptodev that perhaps it too
>>> should be marked experimental and default to off.
>>
>> I think you are right.
>> Declan, what is your opinion?
>
>
> Hey Thomas, yes librte_mbuf_offload should also be set as experimental, it's
> probably one of the areas which will most likely change in the future.
>
> On the issue of turning off experimental libraries in the build by default, my
> preference would be not to turn them off unless the library has external
> dependencies, otherwise the possibility of patches being submitted which
> could break an experimental library will be much higher. In my opinion the
> fewer build configurations developers have to test against the better.

What I'm more worried about is users and developers starting to rely on 
it while still in experimental state, a single comment in the header is 
really easy to miss.

So I'd like to see *some* mechanism which forces users and developers to 
acknowledge the fact that they're dealing with experimental work. 
Defaulting to off is one possibility, another one would be wrapping 
experimental APIs behind a define which you have to set to be able to 
use the API, eg:

#if defined(I_KNOW_THIS_IS_EXPERIMENTAL_AND_MAY_EAT_BABIES)
[...]
#endif

	- Panu -